extract.json · doc-ai v3

Posted to accounts · approved by clerk

Illustrative example. An AI reads a made-up supplier invoice photographed on a clipboard. It reads 6 fields, each with a confidence score from 0.86 to 0.99. The bank account number is hidden before anything is logged. Line items score 0.86, below the 0.90 limit, because a quantity was changed in pencil. A clerk checks them, then the invoice is posted to the accounts system.

Service 04 of 10Technology & Intelligence

AI assistants and automation that take repetitive work off your team.

We build AI for your customers and staff: answers from your documents, chat and voice AI, document reading and task automation. Shuja ur Rahman leads this work.

Typical length
8–14 weeks per feature
Scope
Answers · chat and voice · documents · tasks
Standard
Tested on your examples

04.1What we make/families

4 kinds of AI feature, built to the same 3 rules.

Each is tested on your own examples before launch. You know the running cost from the first test version. When the AI is unsure, a person takes over.

  • A long library aisle lined with shelves of bound volumes

    01

    Knowledge assistants

    Answers from your documents, each showing its source. People see only what they are allowed to open.

    Typical uses

    • Policy and product assistants for staff
    • Help centres that answer questions directly
    • Contract search that shows the clause

    Measured by

    Answered by AI

    Questions solved without a ticket

    Try the answer demo
  • A headset with a microphone rests beside an open laptop on a desk

    02

    Chat and voice AI

    Chat and voice assistants that handle requests in your systems. Hard cases go to a person with all the details.

    Typical uses

    • Order, booking and account support on web and WhatsApp
    • Voice AI that answers calls first
    • Suggested replies and live notes for your staff

    Measured by

    Time per request

    Minutes to solve each request

    Watch a handover
  • Hands measure a machined metal part with a digital caliper on a workshop bench, an inspection sheet beside it

    03

    Documents and images

    AI reads scans, photos and video to pull out data, count items and spot faults. A person checks unsure cases.

    Typical uses

    • Invoices, forms and IDs read into your systems
    • Cap, fill and label checks on a production line
    • Shelf gaps and wrong price labels

    Measured by

    No-touch rate

    Documents posted with no retyping

    Compare before and after
  • Hands typing on a laptop showing a configuration table of workflow rules

    04

    Task automation

    Multi-step tasks across email, documents and your systems. Every step is logged, and a person approves the key ones.

    Typical uses

    • Supplier bills from inbox to accounts
    • New customer sign-up and ID checks
    • Order problems and returns

    Measured by

    Task time

    Hours from start to finish

    Try the automation

04.2Knowledge assistants/inspector

Answers from your documents, checked step by step.

Pick a question about a made-up company. Compare a typical build with ours. Open any step to see what it did. Screens marked Illustrative are examples, not client results.

rag-inspector · your-company-kb · 1,284 docs · 18,402 chunks Illustrative

Questions

text
Sentence not backed by any passage found
text
Backed, but by an old or draft source
1
Link to a passage the model was given

QuestionHow long do customers have to return a damaged item?

A

Typical build

  • Fixed 1,000-token chunks
  • Vector search only
  • No reranker
  • Citations optional

queryHow long do customers have to return a damaged item?

No rewrite, no metadata filter.

  1. 1warranty-terms.pdf§4 Defectsoff-topic0.84Manufacturing defects are covered for 12 months from purchase…
  2. 2returns-policy-v2.pdf§3 Damaged itemsoutdated0.82…damaged items may be returned within 30 days of purchase…
  3. 3returns-policy-v4.pdf§2 · 1,000-token chunkcut mid-table0.79…14 days of delivery … exchanges … gift cards … store credit…
  4. 4shipping-faq.mdDelivery problems0.77If your parcel arrives damaged, keep the packaging and…
  5. 5careers-page.mdBenefitsoff-topic0.71Staff discount on returns and exchanges…

No reranker. All five chunks go to the model in vector-score order, off-topic chunks included.

Context sent to the model≈ 5,000 tokens

Customers can return a damaged item within 30 days of purchase.2outdated source They should keep the original packaging.4 Refunds reach the customer’s account in 3 to 5 working days.not in sources

Shown to the user as written

Scores

Faithfulness0.71
Answer relevance0.83
Context precision0.58
Context recall0.67

A

Model only

  • No documents
  • No sources
  • No citations

queryHow long do customers have to return a damaged item?

Documents switched off. The model answers from what it learned in training, with no access to your files.

Nothing to rank.

Most retailers accept returns of damaged goods within 30 days, and many extend that for defects covered by warranty. Check the retailer’s own policy for details.unsupported · no sources

Plausible, generic and unverifiable

Scores

Faithfulnessn/a
Answer relevance0.61
Context precisionn/a
Context recalln/a

B

Our build

  • Semantic chunks · 300–500 tokens
  • Hybrid BM25 + vector
  • Cross-encoder reranker
  • Citations enforced

rewritedamaged item return window · days from delivery · refund or replacement

filterstatus = current · 1 archived document excluded

  1. 1damaged-goods-sop.docx§1 Intake0.86Support logs the order, the photo and the damage type before…
  2. 2returns-policy-v4.pdf§2.3 Photo evidence0.84Damage reported with a photo within 48 hours skips inspection…
  3. 3returns-policy-v4.pdf§2.1 Damaged on arrival0.83Items that arrive damaged can be returned within 14 days of delivery…
  4. 4shipping-faq.mdDelivery problems0.77If your parcel arrives damaged, keep the packaging and…
  5. 5warranty-terms.pdf§4 Defectsoff-topic0.74Manufacturing defects are covered for 12 months from purchase…

  1. 1returns-policy-v4.pdf§2.1 Damaged on arrivalwas #30.97
  2. 2returns-policy-v4.pdf§2.3 Photo evidencewas #20.91
  3. 3damaged-goods-sop.docx§1 Intakewas #10.62
Context sent to the model≈ 1,100 tokens

Chunks scoring under 0.50 are dropped before generation.

Damaged items can be returned within 14 days of delivery for a replacement or a full refund.1 If the damage is reported with a photo within 48 hours, the inspection step is skipped.2

True to source · every sentence cited · answered

Scores

Faithfulness0.94
Answer relevance0.92
Context precision0.88
Context recall0.90
Test scores for question 1: column A against column B
ScoresAB
Faithfulness0.71n/a0.94
Answer relevance0.830.610.92
Context precision0.58n/a0.88
Context recall0.67n/a0.90

QuestionDoes the Pro plan include single sign-on?

A

Typical build

  • Fixed 1,000-token chunks
  • Vector search only
  • No reranker
  • Citations optional

queryDoes the Pro plan include single sign-on?

No rewrite, no metadata filter.

  1. 1sso-setup-guide.mdConfigure SAML0.89Admins on paid plans can connect an identity provider using SAML 2.0…
  2. 2security-whitepaper.pdf§5 Identityoff-topic0.85Customer data is protected by encryption at rest and SSO support…
  3. 3pricing-plans.mdPlan table · 1,000-token chunkcut mid-table0.80| Feature | Pro | Business | Enter… (row cut at the boundary)
  4. 4release-notes-2026-04.mdSign-in0.78Google and Microsoft sign-in now available on every plan…
  5. 5blog-security.mdAnnouncementoff-topic0.73We believe good security should be available to everyone…

No reranker. All five chunks go to the model in vector-score order, off-topic chunks included.

Context sent to the model≈ 5,000 tokens

Yes, single sign-on is available on every paid plan, including Pro.1not in sources Admins connect an identity provider using SAML 2.0 from the security settings.1

Shown to the user as written

Scores

Faithfulness0.68
Answer relevance0.88
Context precision0.52
Context recall0.50

A

Model only

  • No documents
  • No sources
  • No citations

queryDoes the Pro plan include single sign-on?

Documents switched off. The model answers from what it learned in training, with no access to your files.

Nothing to rank.

Single sign-on is often reserved for business or enterprise tiers, so the Pro plan may not include it. Contact the vendor’s sales team to confirm.unsupported · no sources

Plausible, generic and unverifiable

Scores

Faithfulnessn/a
Answer relevance0.55
Context precisionn/a
Context recalln/a

B

Our build

  • Semantic chunks · 300–500 tokens
  • Hybrid BM25 + vector
  • Cross-encoder reranker
  • Citations enforced

rewritePro plan · SAML single sign-on · included or add-on

filterdoc_type in pricing, docs, release notes

  1. 1sso-setup-guide.mdConfigure SAML0.87Admins on paid plans can connect an identity provider using SAML 2.0…
  2. 2release-notes-2026-06.mdSSO add-on0.85SAML SSO is now available to Pro workspaces as a paid add-on…
  3. 3release-notes-2026-04.mdSign-in0.81Google and Microsoft sign-in now available on every plan…
  4. 4pricing-plans.mdPlans · identity row0.80SAML SSO: Business included · Enterprise included · Pro not included
  5. 5security-whitepaper.pdf§5 Identityoff-topic0.72Customer data is protected by encryption at rest and SSO support…

  1. 1pricing-plans.mdPlans · identity rowwas #40.96
  2. 2release-notes-2026-06.mdSSO add-onwas #20.93
  3. 3sso-setup-guide.mdConfigure SAMLwas #10.71
Context sent to the model≈ 1,100 tokens

Chunks scoring under 0.50 are dropped before generation.

Not by default: SAML single sign-on is included in the Business and Enterprise plans.1 Pro workspaces can add it as a paid add-on, available since June 2026.2

True to source · every sentence cited · answered

Scores

Faithfulness0.96
Answer relevance0.94
Context precision0.91
Context recall0.92
Test scores for question 2: column A against column B
ScoresAB
Faithfulness0.68n/a0.96
Answer relevance0.880.550.94
Context precision0.52n/a0.91
Context recall0.50n/a0.92

QuestionCan I carry unused annual leave into next year?

A

Typical build

  • Fixed 1,000-token chunks
  • Vector search only
  • No reranker
  • Citations optional

queryCan I carry unused annual leave into next year?

No rewrite, no metadata filter.

  1. 1employee-handbook.pdf§7 · 1,000-token chunkcut mid-table0.86…annual leave … public holidays … carried forward … sick leave…
  2. 2holiday-calendar-2026.xlsxSheet 1off-topic0.8126 Jan Republic Day · 15 Aug Independence Day · 2 Oct…
  3. 3leave-policy-2025.pdf§3 Carry-overoutdated0.80Up to 15 days may be carried forward to the next year…
  4. 4payroll-faq.mdLeave encashment0.76Encashment requests are processed with the March payroll…
  5. 5leave-policy-2026.pdf§1 Scope0.74This policy applies to all permanent employees in India…

No reranker. All five chunks go to the model in vector-score order, off-topic chunks included.

Context sent to the model≈ 5,000 tokens

Yes, you can carry forward up to 15 days of annual leave.3outdated source Any days above that are paid out in your final settlement.not in sources

Shown to the user as written

Scores

Faithfulness0.74
Answer relevance0.86
Context precision0.60
Context recall0.55

A

Model only

  • No documents
  • No sources
  • No citations

queryCan I carry unused annual leave into next year?

Documents switched off. The model answers from what it learned in training, with no access to your files.

Nothing to rank.

Carry-over rules vary by employer and by local law. Many companies allow five to ten days to be carried forward. Check your HR policy.unsupported · no sources

Plausible, generic and unverifiable

Scores

Faithfulnessn/a
Answer relevance0.58
Context precisionn/a
Context recalln/a

B

Our build

  • Semantic chunks · 300–500 tokens
  • Hybrid BM25 + vector
  • Cross-encoder reranker
  • Citations enforced

rewritecarry forward unused annual leave · limit · expiry date · exceptions

filtereffective_year = 2026 · 1 superseded policy excluded

  1. 1leave-policy-2026.pdf§1 Scope0.84This policy applies to all permanent employees in India…
  2. 2hr-portal-guide.mdExceptions0.83Managers can approve a carry-over exception under Requests…
  3. 3leave-policy-2026.pdf§3.2 Carry-over0.82Up to 10 unused days carry over and must be used by 31 March…
  4. 4payroll-faq.mdLeave encashment0.75Encashment requests are processed with the March payroll…
  5. 5holiday-calendar-2026.xlsxSheet 1off-topic0.6926 Jan Republic Day · 15 Aug Independence Day · 2 Oct…

  1. 1leave-policy-2026.pdf§3.2 Carry-overwas #30.98
  2. 2hr-portal-guide.mdExceptionswas #20.88
  3. 3leave-policy-2026.pdf§1 Scopewas #10.57
Context sent to the model≈ 1,100 tokens

Chunks scoring under 0.50 are dropped before generation.

Yes, up to 10 unused days carry over into the next year.1 Carried-over days must be used by 31 March, after which they lapse.1 Your manager can approve an exception in the HR portal.2

True to source · every sentence cited · answered

Scores

Faithfulness0.95
Answer relevance0.93
Context precision0.89
Context recall0.94
Test scores for question 3: column A against column B
ScoresAB
Faithfulness0.74n/a0.95
Answer relevance0.860.580.93
Context precision0.60n/a0.89
Context recall0.55n/a0.94

QuestionWhat is the API rate limit on the Team plan?

A

Typical build

  • Fixed 1,000-token chunks
  • Vector search only
  • No reranker
  • Citations optional

queryWhat is the API rate limit on the Team plan?

No rewrite, no metadata filter.

  1. 1rate-limits.mdLimits by plan · 2025outdated0.88Starter 60 · Team 300 · Enterprise custom, requests per minute…
  2. 2api-reference.mdErrors0.84429 Too Many Requests is returned when a limit is exceeded…
  3. 3status-incident-0712.mdPostmortemoff-topic0.80Elevated 429 responses between 14:02 and 14:31 IST…
  4. 4pricing-plans.mdPlan table · 1,000-token chunkcut mid-table0.77| API access | Team | Enterprise | … (row cut at the boundary)
  5. 5sdk-readme.mdRetries0.74The SDK retries idempotent requests with exponential backoff…

No reranker. All five chunks go to the model in vector-score order, off-topic chunks included.

Context sent to the model≈ 5,000 tokens

The Team plan allows 300 requests per minute.1outdated source Requests over the limit return a 429 error.2 Limits reset at the start of every hour.not in sources

Shown to the user as written

Scores

Faithfulness0.70
Answer relevance0.90
Context precision0.62
Context recall0.48

A

Model only

  • No documents
  • No sources
  • No citations

queryWhat is the API rate limit on the Team plan?

Documents switched off. The model answers from what it learned in training, with no access to your files.

Nothing to rank.

API rate limits typically range from 60 to 1,000 requests per minute depending on the plan. The provider’s documentation lists the exact figures.unsupported · no sources

Plausible, generic and unverifiable

Scores

Faithfulnessn/a
Answer relevance0.52
Context precisionn/a
Context recalln/a

B

Our build

  • Semantic chunks · 300–500 tokens
  • Hybrid BM25 + vector
  • Cross-encoder reranker
  • Citations enforced

rewriteTeam plan API rate limit · requests per minute · burst · 429 Retry-After

filterlatest version per document · changelog included

  1. 1api-reference.mdErrors0.86429 Too Many Requests is returned with a Retry-After header…
  2. 2sdk-readme.mdRetries0.80The SDK retries idempotent requests with exponential backoff…
  3. 3rate-limits.mdLimits by plan · 20260.79Team: 600 requests per minute per workspace, bursts to 1,000 over 10 s…
  4. 4changelog-2026-08.md4 August0.78Team rate limit raised from 300 to 600 requests per minute…
  5. 5status-incident-0712.mdPostmortemoff-topic0.66Elevated 429 responses between 14:02 and 14:31 IST…

  1. 1rate-limits.mdLimits by plan · 2026was #30.97
  2. 2changelog-2026-08.md4 Augustwas #40.94
  3. 3api-reference.mdErrorswas #10.90
Context sent to the model≈ 1,100 tokens

Chunks scoring under 0.50 are dropped before generation.

The Team plan allows 600 requests per minute per workspace, raised from 300 on 4 August 2026.12 Short bursts of up to 1,000 requests over 10 seconds are allowed.1 Requests over the limit return HTTP 429 with a Retry-After header.3

True to source · every sentence cited · answered

Scores

Faithfulness0.97
Answer relevance0.95
Context precision0.86
Context recall0.93
Test scores for question 4: column A against column B
ScoresAB
Faithfulness0.70n/a0.97
Answer relevance0.900.520.95
Context precision0.62n/a0.86
Context recall0.48n/a0.93

QuestionWho approves an expense claim over ₹50,000?

A

Typical build

  • Fixed 1,000-token chunks
  • Vector search only
  • No reranker
  • Citations optional

queryWho approves an expense claim over ₹50,000?

No rewrite, no metadata filter.

  1. 1expense-policy-v4-draft.docx§2 Approvals · draftdraft0.87Claims above ₹50,000 need department head sign-off…
  2. 2travel-policy.pdf§6 Hotelsoff-topic0.82Hotel stays above the city cap need prior approval…
  3. 3expense-policy-v3.pdf§2 Approvals0.81Claims above ₹50,000 are approved by the department head…
  4. 4finance-faq.mdReimbursement0.78Approved claims are paid in the next payroll cycle…
  5. 5onboarding-checklist.mdWeek 1off-topic0.72Set up your expense account and corporate card…

No reranker. All five chunks go to the model in vector-score order, off-topic chunks included.

Context sent to the model≈ 5,000 tokens

Expense claims above ₹50,000 are approved by the department head.1draft source Approvals are usually completed within two working days.not in sources

Shown to the user as written

Scores

Faithfulness0.66
Answer relevance0.84
Context precision0.55
Context recall0.46

A

Model only

  • No documents
  • No sources
  • No citations

queryWho approves an expense claim over ₹50,000?

Documents switched off. The model answers from what it learned in training, with no access to your files.

Nothing to rank.

Large expense claims are usually approved by a senior manager or the finance department. Your company’s expense policy will name the approver.unsupported · no sources

Plausible, generic and unverifiable

Scores

Faithfulnessn/a
Answer relevance0.49
Context precisionn/a
Context recalln/a

B

Our build

  • Semantic chunks · 300–500 tokens
  • Hybrid BM25 + vector
  • Cross-encoder reranker
  • Citations enforced

rewriteapprover · expense claim above ₹50,000 · delegation of authority · current

filterstatus = approved · 1 draft excluded

  1. 1finance-faq.mdReimbursement0.83Approved claims are paid in the next payroll cycle…
  2. 2expense-policy-v3.pdf§2 Approvals0.82Claims above ₹50,000 are approved by the department head…
  3. 3delegation-of-authority.xlsxRow 14 · Expenses0.81₹50,000–₹1,00,000: Finance controller · above ₹1,00,000: CFO…
  4. 4travel-policy.pdf§6 Hotelsoff-topic0.70Hotel stays above the city cap need prior approval…
  5. 5onboarding-checklist.mdWeek 1off-topic0.61Set up your expense account and corporate card…

  1. 1delegation-of-authority.xlsxRow 14 · Expenseswas #30.95
  2. 2expense-policy-v3.pdf§2 Approvalswas #20.93
  3. 3finance-faq.mdReimbursementwas #10.41
Context sent to the model≈ 1,100 tokens

Chunks scoring under 0.50 are dropped before generation.

The sources disagree: the delegation-of-authority matrix names the finance controller for claims of ₹50,000 to ₹1,00,000.1 The expense policy (v3) names the department head.2 I have sent this to Finance operations to confirm, with both sources attached.

Sources conflict · confidence 0.52 < 0.70 · handed to Finance operations

Scores

Faithfulness0.93
Answer relevance0.71
Context precision0.84
Context recall0.88
Test scores for question 5: column A against column B
ScoresAB
Faithfulness0.66n/a0.93
Answer relevance0.840.490.71
Context precision0.55n/a0.84
Context recall0.46n/a0.88

The 4 scores

Faithfulness
How much of the answer its sources back up.
Answer relevance
How directly the answer addresses the question asked.
Context precision
Whether the useful passages come first.
Context recall
Whether the search found everything the answer needed.

The names come from free testing tools such as Ragas. We check the AI that scores answers against your experts before trusting it.

Every line needs a source

Every sentence must point to a passage the AI was given, or a check rewrites or removes it. Drafts and old versions stay out of search.

Unsure answers go to a person

When sources disagree or scores are low, the assistant says so and asks the team that owns the answer. Their reply is saved and added to the tests.

04.3Chat and voice AI/conversation

Assistants that know when to hand over.

The assistant handles the requests you allow, using your own systems. When it is unsure or a request needs a person, your team takes over.

Illustrative support conversations. In the chat, a customer reports a cracked blender jug; the assistant looks up the order and books a replacement with confidence 0.93 and 0.88. The customer then reports a duplicate card charge; confidence falls to 0.41, below the 0.60 handover threshold, and refunds need a person, so the assistant hands over to billing support with full context. Priya from billing support reverses the charge. The card number was redacted before logging. On the voice call, a caller changes a delivery address: the caller interrupts the assistant mid-sentence, verifies with a one-time code, and the assistant updates the order with confidence 0.94 and no handover. The phone number, code and address are redacted before logging.

support · conversation 7731
  1. Customer

    My order arrived today but the blender jug is cracked.

  2. Assistant

    Sorry about that. I can see order 48213, delivered this morning. I can send a replacement jug or refund ₹3,499. Which would you prefer?

    orders.lookup · 0.93

  3. Customer

    A replacement, please. Can it arrive before Saturday? It’s a gift.

  4. Assistant

    It can reach you by Friday. I’ve booked the replacement and emailed a free return label for the cracked jug.

    replacements.create · 0.88

  5. Customer

    Also, my card 4111 1111 1111 4242 was charged twice for this order.

  6. Handed to Billing support Confidence 0.41 is under the 0.60 line, and refunds need a person. Transcript, intent, tool calls and a suggested reply attached.
  7. Assistant

    I’ll pass this to our billing team with everything so far, so you won’t need to repeat it. Someone will reply here within a few minutes.

    handover · 0.41

  8. Billing support

    Hi, this is Priya from billing support. I can see the duplicate charge of ₹3,499 and I’ve reversed it. It will show on your statement within 5 to 7 working days.

    Sent by Priya · Billing support

Voice call · order support

Streaming · barge-in on · 00:19

  1. 00:00.8 CallerHi, I need to change the delivery address for my order.
  2. 00:02.9 AssistantOf course. Is that order 48213, arriving on Friday?
  3. 00:05.1 Caller · barge-inYes, that one. Send it to my office instead, the address on my account.
  4. 00:05.2 Barge-in detected · assistant audio stopped in 90 ms
  5. 00:08.4 AssistantI’ve sent a one-time code to the number ending 42. Could you read it out?
  6. 00:14.7 CallerIt’s [code redacted].
  7. 00:16.2 AssistantThank you, that’s verified. Your order will now go to your office address on Friday.
  8. 00:19.0 Resolved · no handover · call 00:19

Time to first audio Illustrative

  • Speech-to-text 190 msWhisper large-v3 turbo, streamed
  • AI reply starts 310 mssmall model, routed
  • Text-to-speech first audio 220 msElevenLabs Flash
  • Total 720 mstarget under 800 ms
  • Live voice calls

    Speech becomes text, the AI replies and the reply is spoken, all in one flow. Callers can cut in at any time. We aim for the first words within about 0.8 seconds.

  • Handover with context

    For each type of request, we set what the AI may do alone. Some topics always go to a person. Whoever takes over sees the chat, what was done and a suggested reply.

  • Private details hidden

    Card and phone numbers, one-time codes and IDs are hidden before a chat is saved, analysed or reused for tests.

04.4Documents and images/vision

AI for images that counts, reads and checks.

Drag the handle to see what the AI saw. Each item it finds gets a label and a confidence score. Unsure ones go to a person.

Illustrative detectionsReview threshold 0.70

7 dark wine bottles with yellow capsules pass along a conveyor in front of stainless steel tanks
Line 2 · camera 3 · frame 18,204detector v2.3 · edge GPU · 14 ms

7 bottles · 1 spacing fault · 1 to re-check

Bottling line: detections per class with precision and recall on the validation set, and the action each one triggers (illustrative)
ClassCountPR
capsule_okPass60.990.98
pitch_shortSlow the line to space them10.950.93
occludedRe-check at camera 4 · 0.62 < 0.7010.880.90
Precision
Of everything flagged, the share that was right. Low precision wastes people’s time.
Recall
Of all defects present, the share that was caught. Low recall lets problems through.
A chilled supermarket shelf of milk and soy milk cartons in rows above electronic price labels
Store 114 · aisle 6 · photo 09:12detector v1.8 + OCR · phone app · 0.9 s

17 products in view · 1 wrong price · 1 gap

Retail shelf: detections per class with precision and recall on the validation set, and the action each one triggers (illustrative)
ClassCountPR
facingsShelf plan check170.970.95
price_tagRead against the price list90.980.96
price_mismatchTask to store staff10.910.88
gapRestock request10.940.92
low_stockHuman review · 0.66 < 0.7010.790.83
Precision
Of everything flagged, the share that was right. Low precision wastes people’s time.
Recall
Of all defects present, the share that was caught. Low recall lets problems through.

Route A

A detector trained on your images

A detector such as YOLO learns from your labelled pictures. At the camera, it checks each frame in under a tenth of a second. It suits counting and fast production lines.

Frames per secondPrecision and recall per item typeWorks offline

Route B

AI model that reads documents

The model reads layouts, tables, stamps and handwriting into clean data. Every field is checked before anything posts. Each page takes seconds, in your cloud or with a provider under a data agreement.

Accuracy per fieldNo-touch rateField checks

Both routes

People check the unsure cases

Reviewers confirm or fix unsure results in one click, and each fix becomes training data. A new model goes live only if it beats the old one on unseen examples.

Review queueTraining dataUnseen test examples

04.5Task automation/automation

Invoices from inbox to accounts, with a person to approve.

An invoice arrives by email and is posted to your accounts system (ERP). Failed steps retry without posting twice. Anything over the spending limit waits for a person. Try approving one.

A thick, uneven stack of printed invoices, receipts and forms
The work it removesInvoices, receipts and forms that someone types up today become clean records.Photo: Camilo Rueda Lopez / Unsplash
accounts-payable · workflow v14 · survives restarts r-2291 · completed Illustrative
Accounts payable workflow: email trigger, classify, extract, validate with retries, branch on amount, human approval over ₹2,00,000, post to ERP, notify the team, audit record and failed runs set aside after three retries. yes · over limit no after 3 failed retries retry · backoff 2 s, 4 s, 8 s TRIGGER Email with PDF ap@your-company · 1 PDF CLASSIFY AI · document type invoice · 0.98 EXTRACT Vision · 12 fields min confidence 0.95 VALIDATE Rules + ERP lookup PO matched · 3-way ok BRANCH Over ₹2,00,000? ₹3,40,000 · yes APPROVAL Finance lead approved · 3 m 02 s POST ERP document 5100042 · idempotent NOTIFY Team channel #ap-approvals · sent AUDIT Run record r-2291 · 14 events FAILED RUNS After 3 retries empty
run console · r-2291
  1. triggeremail received · ap@your-company · inv-88412.pdf
  2. classifyinvoice · 0.98 · small model
  3. extract12 fields · min confidence 0.95
  4. validateERP lookup timed out · retry 1 in 2 s
  5. validatePO 4500123 matched · 3-way match ok
  6. branch₹3,40,000 over ₹2,00,000 · approval required
  7. approvalapproved by Finance lead
  8. postERP document 5100042 · key inv-88412
  9. notify#ap-approvals · message sent
  10. auditrun r-2291 closed · 14 events

Approval step · ₹3,40,000 · approved by Finance lead

  • 78%No-touch runs this week
  • 41 sMedian cycle, no approval
  • 6 · 1Retries · failed runs
Recent workflow runs with amount, path, duration, retries and status (illustrative)
RunAmountPathDurationRetriesStatus
r-2291inv-88412.pdf₹3,40,000Approval3 m 08 s1Posted · approved
r-2290inv-88409.pdf₹48,200No-touch posting38 s0Posted
r-2289inv-88401.pdf₹1,12,000No-touch posting41 s0Posted
r-2288inv-88397.pdf₹2,75,000Approval—0Waiting for approval
r-2287inv-88390.pdf—Failed run2 m 05 s3ERP unavailable · replayed
  • No double posting

    Each entry has a unique key, so a retry never posts twice.
  • Automatic retries

    Temporary errors retry after 2, 4 and 8 seconds, then the run fails. Lasting errors stop at once.
  • Failed runs

    Runs that use up their retries are kept with their data and error, ready to rerun after a fix.
  • A record of every run

    Each run records who approved what, which model version decided, and every input and output.

Runs on tools such as

  • n8n
  • Temporal
  • SAP
  • Python
  • Zapier

AlsoPower AutomateSlackMicrosoft TeamsGmail

04.6How it works/stack

5 steps from documents to answers, each easy to swap.

Each step has a job, a reason and our preferred tools. We build the links between steps, so better models or tools can be swapped in.

pipeline · knowledge-assistant · v24 · hosted + self-hosted
  1. 01

    Read the documents

    Documents, tickets, tables and scans become clean text, tagged with source, version, owner and who may read it. Bad reading is the top cause of wrong answers.

    We work with

    • Python either
    • Apache Airflow either
    • Airbyte either
    • Apache Kafka either
    • PostgreSQL either

    AlsoUnstructured

    1,284 docs · 18,402 chunks · 300–500 tokens

  2. 02

    Index by meaning

    Each passage is filed by meaning and by keyword, with who may see it. Search uses both and never shows what a user could not open.

    We work with

    • PostgreSQL + pgvector either
    • Qdrant either
    • Elasticsearch either
    • Milvus either
    • Hugging Face either

    AlsoWeaviatePinecone

    1,024-d · HNSW · nightly re-index

  3. 03

    Search & rank

    Keyword and meaning search find 20 likely passages. A second check keeps the best 3, so answers are better and cheaper.

    We work with

    • LangChain either
    • Elasticsearch either
    • Redis either

    AlsoLlamaIndexCohere Rerank

    hybrid · top 20 → rerank 3 · 41 ms

  4. 04

    Write & check

    Small jobs go to a small model and drafting to a larger one. Checks block hidden instructions meant to trick the AI. They also check each answer for format, sources and personal data.

    We work with

    • Anthropic hosted API
    • Google Gemini hosted API
    • Mistral AI either
    • Meta Llama self-hosted
    • DeepSeek either
    • vLLM self-hosted
    • LangGraph either

    AlsoOpenAIGuardrails

    Voice in and out

    Whisper (open weights)DeepgramElevenLabsAzure AI Speech

    hosted models · 95% under 1.8 s · streamed

  5. 05

    Test & watch

    Every release is tested on your own examples first. Each live answer is logged with its model, sources, cost and speed. A sample is scored every day.

    We work with

    • MLflow either
    • OpenTelemetry either
    • Grafana either
    • Prometheus either
    • GitHub Actions either

    AlsoRagasLangfuse

    24 auto-checks · daily sample of 200

Every swap is tested

Swapping a model or any part is a settings change and a test run on your examples. New models come out monthly, and you can use one that week.

Open models on your servers

Some data must stay in-house or in India. Then Llama, Mistral or DeepSeek models run in your private cloud or an Indian data centre. Nothing leaves your network, and the same tests apply.

Chosen on test scores

We score each option on your own examples for quality, speed and cost per answer. You keep the comparison, and we re-run it when a new option appears.

04.7Release checks/quality

Quality checks every change must pass before users see it.

Every change to an AI instruction, model or document is a release, tested on examples your experts approved. Pick a check to see what it tests.

release · assistant v24 · prompt v23 → v24 · model upgrade Released · 5% of users → 100% Illustrative

8 checks · 1 blocked then fixed · 1 human approval12 m 40 s

Check 01 of 8

Expert answer tests

Every question in your test set is answered by the new version. Each answer is scored for faithfulness, relevance, context precision and context recall. The scoring AI is first checked against your experts. A score under the limit blocks the release. Failing answers are listed with the passages they used, so the fix is clear.

Catches

  • Worse answers after an AI change
  • Worse search after a document update

Runs with

  • Ragas
  • Langfuse
  • GitHub Actions

Check 02 of 8

Live question replay

Last month’s live questions, with names removed, are run again through the new version. We list any sentence the source check now flags as unsupported. We also list any answer that now goes to a person without a reason.

Catches

  • Silent behaviour change
  • Handover rate creeping up

Runs with

  • Replay harness
  • Langfuse

Check 03 of 8

Attack tests

Attack prompts test 5 risks from the OWASP list for AI. They try hidden instructions (LLM01) and attempts to pull out other users’ data or secrets (LLM02). They try output that is unsafe when used elsewhere (LLM05) and leaks of the AI’s own instructions (LLM07). They also try confident answers where the sources are silent (LLM09). Every attack must fail. An attack that once worked stays in the tests for good.

Catches

  • LLM01 prompt injection
  • LLM02 sensitive information disclosure
  • LLM05 improper output handling
  • LLM07 system prompt leakage
  • LLM09 misinformation

Runs with

  • Promptfoo
  • Garak
  • Custom suites

Check 04 of 8

Private data tests

We plant fake card numbers, phone numbers and ID numbers in documents and user messages. The release passes only if none show up in answers, logs, analytics or test data. This proves private details are hidden before anything is saved.

Catches

  • Personal data in logs
  • Personal data in test sets

Runs with

  • Presidio
  • Regex + checksum
  • Trace assertions

Check 05 of 8

Speed and cost limits

The test run also checks speed and cost. We measure the slowest 5% of replies (p95) and the cost per answer. Both are compared with the limits agreed at the test-version stage. A better model at 3 times the cost is a talk with you, not a surprise on the invoice.

Catches

  • Cost creeping up
  • Replies slowing down

Runs with

  • OpenTelemetry
  • Grafana
  • Cost model

Check 06 of 8

Accessibility of AI screens

Screen readers announce each answer when it is complete, not word by word. Source links have clear names, the stop button works by keyboard, and contrast holds in light and dark mode. Automatic checks run on every change, and a person tests each major release.

Catches

  • WCAG 2.2 AA
  • Screen-reader announcements

Runs with

  • axe-core
  • Playwright
  • Manual pass

Check 07 of 8

Content credentials

When a feature makes images, audio or video, we attach a C2PA content credential as it is made. It states that AI made the media, and which model. This check confirms the record survives every step, so the person who sees the media is told.

Catches

  • Undisclosed synthetic media
  • Manifest stripped by resizing

Runs with

  • c2pa-rs
  • Signing key in KMS

Check 08 of 8

Human sign-off

A person accountable for the feature reads the test summary, the flagged answers and the attack report. They approve a release to 5% of users. It widens over 24 hours only if live answers keep to their sources. The approval, the evidence and the rollout form one record.

Catches

  • Unowned releases
  • All-at-once releases

Runs with

  • Release record
  • Feature flags

Standards we build to

We drew these badges for the standards our checks follow. They show how we work. We hold no certificates for them.

  • OWASP Top 10 for LLM ApplicationsSecurity risks in generative AI applicationsRisks in AI apps, such as hidden instructions that trick the AI, data leaks and AI with too much power.
  • ISO/IEC 42001:2023Artificial intelligence management systemsRules for managing AI in a company: policy, impact checks and controls on data and models.
  • NIST AI RMF 1.0AI Risk Management FrameworkManaging AI risk in 4 parts: Govern, Map, Measure and Manage. A second guide covers generative AI (NIST AI 600-1).
  • WCAG 2.2 AAWeb Content Accessibility GuidelinesRules that make websites usable by people with disabilities. Version 2.2 adds visible focus, tap size and easier log-ins.
  • GDPRGeneral Data Protection Regulation (EU) 2016/679Europe's data law: a legal reason to use data, people's rights, privacy by design, breach reports and risk checks.
  • DPDP Act 2023Digital Personal Data Protection Act, 2023India's data law and its Rules: notice and consent, people's rights, breach reports and extra duties for significant firms.
  • EU AI ActArtificial Intelligence Act (EU) 2024/1689Europe's AI law sets rules by risk: banned uses, duties for high-risk AI and general-purpose models, and AI labels.
  • C2PA content credentialsOrigin records for AI mediaAn open standard that attaches a signed record to images, audio and video. It shows whether AI made or edited them.

04.8Cost & energy/efficiency

Lower cost and energy per answer.

The same question can cost many times more, depending on design. We measure cost and energy per answer, then cut both without lowering test scores.

cost lab · knowledge-assistant · per 1,000 answers Illustrative

Cost per 1,000 answers₹90

  • Retrieval & rerank ₹28
  • Small model ₹20
  • Large model ₹36
  • Cache ₹6

Energy per 1,000 answers150 Wh

  • Retrieval & rerank 30 Wh
  • Small model 40 Wh
  • Large model 75 Wh
  • Cache 5 Wh

As B, plus a semantic cache that serves the thirty-five percent of questions asked before, and prompt caching for the static system prompt and tool schemas. ₹90 per 1,000 answers is ₹0.09 each, inside the ₹0.12 release budget.

Text sent to the model per answer
940
p95 latency
1.2 s
Monthly cost at volume
₹18,000
Monthly energy at volume
30 kWh
Test-set faithfulness
0.93 A 0.90 · B 0.93 · C 0.93 · all clear the 0.90 bar
Against design A
−91% cost · −85% energy

Saved answers

Repeat questions, even in other words, get a saved, checked answer. Saved answers keep their sources and expire when those sources change.

35% of traffic · 0 model tokens

Saved instructions

The instructions and policy text are the same on every call. AI providers and vLLM can store that fixed part, which cuts most of its cost.

static prefix · up to 90% off input

Less text for the model

We keep only the best 3 passages out of 20. The model reads a fifth of the text and answers more precisely.

5,000 → 1,100 tokens

Small models for the small jobs

Small models sort each question by topic and language, and spot personal data. The large model only gets the questions that need it.

70% routed small

Big jobs in low-carbon hours

Re-indexing and nightly tests run when a forecast says the local grid is cleanest. Off-peak rates cut the cost too.

carbon-aware window

SCI · ISO/IEC 21031:2024Software Carbon IntensityHow we apply it We measure carbon per request, user or AI answer as a starting point.

SCI = ((E × I) + M) per R Energy used, times the carbon in that power, plus the carbon of making the hardware, per unit of work. Here R is one answer.

04.9Plan and timing/process

From test version to launch in 4 stages.

Before we build, we agree with your experts what a good answer looks like. We score every change against it. You see a working test version on your own documents by week 3. After launch, we release changes weekly.

delivery plan · one AI feature · 14 weeks · 4 checkpoints Typical plan

A typical 14-week plan. Plan runs weeks 1 to 2 and Test version weeks 2 to 5. Build runs weeks 5 to 12 and Launch & run weeks 12 to 14. Each ends with a checkpoint. The test set grows from 40 questions in week 2 to 400 by week 14. From week 4, every build is scored against it. One release in week 9 is blocked and fixed the next day. Weekly live releases start in week 12.

Phase 01 of 4Wk 01–02

Plan

We agree the task, users, data and the mistakes that matter. With your experts, we draft measures and a test set.

  • Feature plan
  • Test set v1
  • Data list
Checkpoint G1
Your experts approve the first test set: 40 to 60 questions or documents from your work, with answers and sources.
Every week you see
2 sessions with the people who own the answers. A list of your data and who may access it by the end of week 2.
Where AI helps
AI groups past tickets and searches into draft questions. Your experts review the draft instead of starting from nothing.

Phase 02 of 4Wk 02–05

Test version

We compare search, instructions and models on your test set for quality, speed and cost. The best scores win.

  • Test version
  • Model comparison
  • Cost estimate
Checkpoint G2
The chosen design beats a basic build on your test set, within the speed and cost limits agreed for launch.
Every week you see
A working test version on your own documents. The model comparison is updated every Friday.
Where AI helps
AI scores 3 or 4 designs overnight on your test set, so we choose on numbers within days.

Phase 03 of 4Wk 05–12

Build

We build the live system, checks, screens, feedback and monitoring. Every release must pass the test set.

  • Live feature
  • Automatic checks
  • Tests on every release
Checkpoint G3
All 8 release checks pass, and a small group of named users has used it daily for 2 weeks.
Every week you see
A demo every Friday on a test site, with that week’s test scores. We only show work that passed the checks.
Where AI helps
AI drafts code and tests. A named engineer reviews every change, and it is scored again before it goes in.

Phase 04 of 4Wk 12–14

Launch & run

A staged launch, live scoring and alerts when quality drops or costs rise. The test set grows from live use.

  • Rollout
  • Quality dashboard
  • List of fixes
Checkpoint G4
Live answers stick to sources and handovers hold steady for 4 weeks. Then your team takes over, with guides and dashboards.
Every week you see
Weekly releases that each pass the tests, and a monthly quality and cost review with the owner.
Where AI helps
AI scores a sample of live answers every day, and a person reviews the flagged ones.

04.10What you get/deliver

Code, AI instructions and tests your team can run without us.

You get the code, AI instructions, test sets, dashboards and guides in one handover. Open any file.

Open · ready to mergehandover/v24 → mainyour-company/ai-assistant

Handover #1: assistant v24 and accounts-payable automation

9 deliverables141 files changed6 of 6 checks passed2 of 2 approvals from your team

  1. app/The AI feature, in production Source 64 files

    API, streaming UI with citations and the handover path, deployed to your cloud.SourceTypeScriptPython

    app/answer/answer.ts+11 · excerptTS

    1. // Answer with sources, or hand over. Never an uncited guess.
    2. export async function answer(q: Question, user: User) {
    3. const found = await retrieve(q, { acl: user.groups, topK: 20 });
    4. const top = await rerank(q, found, { keep: 3, minScore: 0.5 });
    5. if (top.length === 0) return handover(q, "no-sources");
    6. ​
    7. const draft = await generate(q, top, { model: route(q) });
    8. const checked = enforceCitations(draft, top);
    9. if (checked.confidence < 0.7) return handover(q, "low-confidence", checked);
    10. ​
    11. trace.log({ q: redact(q), sources: top.map((s) => s.id), cost: draft.cost });
    12. return checked;
    13. }

    e41c09afeat: hand over when nothing passes the reranker2 days ago

  2. pipelines/ingest/Ingestion pipelines & index Code 18 files

    Parse, chunk, embed and index, nightly and on change, with permissions carried through.CodeAirflow DAGVector index

    pipelines/ingest/dag.py+10 · excerptPY

    1. # Nightly: parse, chunk, embed and index new or changed documents.
    2. @dag(schedule="0 2 * * *", start_date=datetime(2026, 1, 1), catchup=False)
    3. def ingest_knowledge_base():
    4. docs = changed_since_last_run(sources=["sharepoint", "confluence", "drive"])
    5. parsed = parse(docs, keep_tables=True, ocr="scanned-only")
    6. chunks = semantic_chunk(parsed, min_tokens=300, max_tokens=500, overlap=80)
    7. tagged = attach_meta(chunks, fields=["source", "version", "owner", "acl"])
    8. index.upsert(embed(tagged), keyword_index=True)
    9. retire(superseded(docs)) # old versions leave the index, not only the UI
    10. ​
    11. ingest_knowledge_base()

    7b2d5f0fix: retire superseded versions from the indexlast week

  3. config/Prompt & model registry YAML 9 files

    Models, routes, retrieval settings and prompts under version control. Every change is a release.YAMLVersioned

    config/assistant.yaml+13 · excerptYAML

    1. # Every change here is a release: it runs the eval gates in CI.
    2. version: 24
    3. router:
    4. classify: small-model # intent, language, PII, routing
    5. answer: large-model # drafting and reasoning
    6. fallback: open-weights-70b # served with vLLM in your VPC
    7. retrieval:
    8. search: hybrid # BM25 + vector
    9. candidates: 20
    10. rerank: { model: cross-encoder, keep: 3, min_score: 0.50 }
    11. prompts:
    12. system: prompts/assistant.v24.md
    13. handover: prompts/handover.v7.md

    c90a3e1chore: v24 routes answers to large-model, open-weights fallbackyesterday

  4. evals/Evaluation suite Golden set 12 files

    The golden set your experts signed, the calibrated judge and the thresholds that block a release.Golden setScorersCI gate

    evals/gates.yaml+12 · excerptYAML

    1. golden_set: evals/golden-set.jsonl # 400 questions · owner: Support ops
    2. judge: calibrated-judge-v3 # agreement with expert labels: 0.91
    3. gates:
    4. faithfulness: { min: 0.90 }
    5. answer_relevance: { min: 0.85 }
    6. context_precision: { min: 0.80 }
    7. context_recall: { min: 0.80 }
    8. red_team: { successful_attacks: 0 } # OWASP LLM01, 02, 05, 07, 09
    9. pii_leakage: { found: 0 }
    10. p95_latency_s: { max: 2.5 }
    11. cost_per_answer_inr: { max: 0.12 }
    12. on_fail: block_release

    3f8b6d2test: add 40 billing questions to the golden setyesterday

  5. guardrails/Guardrails & review queues Policy 7 files

    Input and output checks, redaction and the rules for when a person takes over.PolicyWorkflow

    guardrails/policy.yaml+11 · excerptYAML

    1. input:
    2. prompt_injection: { detector: classifier, action: refuse_and_log } # LLM01
    3. pii: { detect: [card, aadhaar, phone, email], action: redact_before_log }
    4. output:
    5. citations: required # unsupported sentences are removed
    6. schema: answer.schema.json # tool calls and JSON validated
    7. render: plain_text # no raw HTML or links from the model (LLM05)
    8. handover:
    9. when: [confidence_below_0.70, sources_conflict, topic_in_refunds_or_legal]
    10. queue: support-tier-2
    11. attach: [transcript, intent, tool_calls, sources, suggested_reply]

    a17e4c8feat: redact Aadhaar numbers before logging3 days ago

  6. workflows/Automation workflows Temporal 23 files

    Durable, idempotent workflows with retries, approval gates and a dead-letter queue.Temporaln8n

    workflows/invoice_to_erp.py+18 · excerptPY

    1. # Retries after 2 s, 4 s and 8 s; a fourth failure goes to the dead-letter queue.
    2. RETRY = RetryPolicy(initial_interval=timedelta(seconds=2), backoff_coefficient=2.0, maximum_attempts=4)
    3. ​
    4. @workflow.defn
    5. class InvoiceToERP:
    6. def __init__(self) -> None:
    7. self.decision: str | None = None
    8. ​
    9. @workflow.signal
    10. def decide(self, decision: str) -> None: # "approve" or "reject", from the finance lead
    11. self.decision = decision
    12. ​
    13. @workflow.run
    14. async def run(self, email_id: str) -> str:
    15. doc = await workflow.execute_activity(extract_invoice, email_id, start_to_close_timeout=timedelta(seconds=60), retry_policy=RETRY)
    16. await workflow.execute_activity(validate_against_erp, doc, start_to_close_timeout=timedelta(seconds=30), retry_policy=RETRY)
    17. if doc.total > APPROVAL_LIMIT:
    18. await workflow.wait_condition(lambda: self.decision is not None)
    19. if self.decision == "reject":
    20. return await workflow.execute_activity(notify_rejected, doc, start_to_close_timeout=timedelta(seconds=30))
    21. return await workflow.execute_activity(post_to_erp, doc, start_to_close_timeout=timedelta(seconds=30), retry_policy=RETRY)

    5d0c9b7fix: idempotency key on the ERP post steplast week

  7. dashboards/Quality, latency & cost dashboards Grafana 6 files

    The numbers the feature is judged on, with alerts wired to an owner.GrafanaLangfuse

    dashboards/assistant.json+11 · excerptJSON

    1. {
    2. "title": "Knowledge assistant · quality, latency, cost",
    3. "panels": [
    4. { "title": "Groundedness, daily sample", "threshold": 0.90 },
    5. { "title": "Handover rate by intent", "unit": "percent" },
    6. { "title": "p95 latency by model route", "unit": "s" },
    7. { "title": "Cost per answer", "unit": "INR" },
    8. { "title": "Top unanswered questions", "type": "table" }
    9. ],
    10. "alerts": ["groundedness_7d < 0.90", "cost_per_answer > budget"]
    11. }

    b62f1a4feat: alert when 7-day groundedness is under 0.904 days ago

  8. docs/runbook.mdRunbooks Markdown 1 file

    What to do when a number moves, written so your team can act without us.MarkdownOn-call

    docs/runbook.md+8 · excerptMD

    1. # Runbook · knowledge assistant
    2. ​
    3. ## Groundedness below 0.90 for a day
    4. 1. Open the flagged answers in Langfuse and group them by source.
    5. 2. Check index freshness: a failed nightly ingest is the usual cause.
    6. 3. Re-run ingestion for that source, then re-score the sample.
    7. 4. If it persists, roll config back to the last good version.
    8. ​
    9. ## Provider outage or rate limits
    10. Traffic fails over to the fallback route. Answers stay cited.

    91e7c3ddocs: provider outage and rate-limit stepslast week

  9. docs/model-card.mdModel cards Markdown 1 file

    Intended use, limits, evaluation and owner for each release, aligned with ISO/IEC 42001 documentation.MarkdownPer release

    docs/model-card.md+8 · excerptMD

    1. # Model card · assistant v24
    2. ​
    3. Intended use Staff questions answered from approved policy documents
    4. Out of scope Legal or medical advice; decisions about individuals
    5. Models small-model (routing) · large-model (answers)
    6. Data 1,284 documents, permission-filtered at query time
    7. Evaluation Golden set of 400 · faithfulness 0.93 · relevance 0.94
    8. Known limits Tables in scanned PDFs; questions spanning 5+ documents
    9. Owner Head of Support operations · reviewed each release

    0c4a8f6docs: model card for assistant v24yesterday

04.11What we measure/outcomes

Agreed measures that decide if the AI stays on.

Most AI trials stall because nobody owns the result. We agree the owner and the measure before we build. We also set the level that switches the AI off.

production · assistant and invoice automation · 12 weeks after launch liveIllustrative
  • Answered by AI

    41%+23 pts

    was 18% via searchtarget ≥ 40%

    Questions resolved with no ticket. Twelve weeks: 22, 26, 29, 31, 33, 35, 36, 37, 38, 39, 40, 41.

  • Time per request

    6.1 min−35%

    was 9.4 mintarget ≤ 7 min

    Handed-over conversations. Twelve weeks: 9.1, 8.6, 8.2, 7.8, 7.5, 7.1, 6.9, 6.7, 6.5, 6.3, 6.2, 6.1.

  • No-touch rate · invoices

    78%+78 pts

    was 0%, all manualtarget ≥ 75%

    Invoices posted with no touch. Twelve weeks: 52, 58, 63, 66, 69, 71, 72, 74, 75, 76, 77, 78.

  • Cost per solved case

    ₹0.55−50%

    was ₹1.10 in week 1budget ≤ ₹0.70

    Model and infrastructure, about six answers at ₹0.09. Twelve weeks: 1.1, 1.02, 0.95, 0.84, 0.76, 0.72, 0.7, 0.64, 0.6, 0.58, 0.56, 0.55.

  • True to source

    0.96+0.05

    floor 0.90target ≥ 0.93

    Daily sample of 200 answers. Twelve weeks: 0.91, 0.92, 0.93, 0.93, 0.94, 0.93, 0.92, 0.93, 0.94, 0.95, 0.95, 0.96.

Answers you can check

Each answer shows its sources and gets an accuracy score.

True to source · weekly average of daily checks0.96
0.80 0.85 0.90 0.95 1.00 1 2 3 4 5 6 7 8 9 10 11 12 floor 0.90 W7

Week 7: a policy upload failed to index and answers cited a stale version. One day’s sample fell to 0.87 (daily low; weekly average 0.92). The alert fired the same day; fixed in 26 hours.

Every answer carries its sources; 0 uncited answers shown to users

Hard cases go to people

Unusual cases reach a person with all the details. We automate more as accuracy is shown.

No-touch rate · invoices posted with no human step78%
40% 50% 60% 70% 80% 1 2 3 4 5 6 7 8 9 10 11 12 target 75% W6

Week 6: the review threshold moved from 0.95 to 0.92 after 1,200 checked cases showed no loss in accuracy.

Exceptions reach a named person with the document, the fields and the reason

Running costs known up front

We track cost per request and per finished task from the first test version.

Cost per solved case · model and servers₹0.55
₹0.40 ₹0.60 ₹0.80 ₹1.00 ₹1.20 1 2 3 4 5 6 7 8 9 10 11 12 budget ₹0.70 W4

Week 4: saved answers went live and now cover about a third of questions with no model call.

Cost per answer tracked from the test version, not found on the first invoice

Offers & plans

AI features your customers use, and automation your team can trust.

From knowledge assistants to automation for critical work, each offer lists what is included, typical timing and who it suits.

Categories
05
Offers
17
Plans
04
Not sure what you need? Tell us the problem

How to buy

  1. 01Pick offers. Ask about one, or add several to one enquiry.
  2. 02Choose a plan. A small job, a fixed project or ongoing help.
  3. 03Send your enquiry. We reply within one working day.

Browse by category

Timelines are typical. We write the scope before we quote.

02Chat and voice AI

4 offers
Typical timeline: 6–10 weeks

Customer support assistant

A chat assistant for your site or app that handles orders, returns and bookings. Harder cases go to a person with full details.

What’s included

  • Topics drawn from your support tickets
  • Linked to order, customer or booking systems
  • Handover to your helpdesk with context
  • Solved, handover and satisfaction rates
  • Chat
  • Handover
  • Solved rate
  • OpenAI

Best forSupport teams answering lots of repeat questions.

Typical timeline: 4–8 weeks

WhatsApp & messaging assistant

An assistant on WhatsApp and other chat apps that handles enquiries, bookings, reminders and updates.

What’s included

  • WhatsApp Business set-up and templates sent to Meta
  • Chat flows, with AI for other questions
  • Opt-in and consent handling
  • Every conversation logged in your CRM
  • WhatsApp Business
  • Templates
  • Opt-in
  • Twilio

Best forBusinesses in markets such as India where customers prefer WhatsApp.

Typical timeline: 6–12 weeks

Voice AI agent

A phone assistant that understands everyday speech, books appointments and sends calls to the right place or person.

What’s included

  • Speech recognition and a natural voice
  • Call flows linked to your phones
  • Transfer to staff with a summary
  • Recordings and transcripts, with caller consent
  • Voice
  • Phone lines
  • Call summaries
  • Twilio
  • OpenAI

Best forClinics, service firms and call centres with many calls.

Typical timeline: 4–8 weeks

Sales enquiry assistant

An assistant that answers product questions, sorts new enquiries and books meetings for your sales team, day or night.

What’s included

  • Screening questions from your sales process
  • Product answers from approved content only
  • Calendar booking
  • Customer record with a chat summary
  • Enquiry sorting
  • Booking
  • CRM
  • Salesforce
  • OpenAI

Best forSales teams losing enquiries that arrive after hours.

03Documents and images

3 offers
Typical timeline: 6–10 weeks

Document reading

AI reads invoices, forms, contracts and IDs and pulls out the fields you need. A person checks unsure cases.

What’s included

  • Reading text and key fields
  • Sorting by document type
  • Unsure cases sent for review
  • Export to your accounts or sales system
  • Reading
  • Data capture
  • Person checks
  • AWS

Best forFinance, insurance, shipping and sign-up teams typing data by hand.

Typical timeline: 8–16 weeks

Camera inspection

Cameras and AI that spot faults, count items or check rules in images and video.

What’s included

  • Image set and labelling plan
  • Model training and testing
  • On-site or cloud set-up
  • Alerts and a review dashboard
  • Camera AI
  • On-site set-up

Best forManufacturing, retail and field operations with visual checks.

Typical timeline: 6–10 weeks

Contract review assistant

It summarises long documents, compares clauses with your standard terms and flags what a lawyer should see.

What’s included

  • Clause library and review guide
  • Clause finding and comparison
  • Risk flags linked to the text
  • Steps for reviewers
  • Summaries
  • Clause comparison
  • Reviewer-led
  • OpenAI
  • LlamaIndex

Best forLegal, buying and compliance teams reviewing many documents.

04Task automation

3 offers
Typical timeline: 2–6 weeks

No-code automation

We link email, forms, customer records and spreadsheets with n8n, Make or Zapier, so routine steps run by themselves.

What’s included

  • Task map and automation shortlist
  • Automations with error alerts
  • AI to sort, draft and summarise
  • Guides and team handover
  • n8n
  • Make
  • Zapier

Best forSmall and mid-size teams losing hours to copy-and-paste work.

Typical timeline: 8–14 weeks

Automation for critical work

Custom automation for important, high-volume work. It recovers from errors, retries safely and records every step. Built on Temporal or similar.

What’s included

  • Steps designed, problems sent to people
  • Recovery with retries and time limits
  • Steps a named person approves
  • Monitoring and a full record
  • Temporal
  • Error recovery
  • Problems to people

Best forWork where a dropped step costs money: payments, claims, orders.

Typical timeline: 4–8 weeks

Inbox sorting

AI reads emails and tickets, sorts them by topic and urgency, drafts replies and sends each to the right team.

What’s included

  • Sorting by topic, urgency and mood
  • Draft replies a person reviews
  • Sending rules for your helpdesk
  • Accuracy reports
  • Sorting
  • Draft replies
  • Right team
  • OpenAI

Best forShared inboxes and help desks with slow first replies.

05AI in your product & testing

4 offers
Typical timeline: 6–12 weeks

AI features in your product

Summaries, drafts, suggestions and plain-language search built into your product, with a backup when the AI is unsure.

What’s included

  • Feature ideas with your product team
  • Screens for AI answers and feedback
  • Model set-up with cost limits
  • Test set and release check
  • In-product AI
  • Backups
  • Cost per request
  • OpenAI

Best forSoftware products adding AI features their customers will use.

Typical timeline: 8–12 weeks

Personal recommendations

Suggest the next product, article or action for each user, based on what they and similar users do.

What’s included

  • Data review and design
  • Recommendation models
  • A/B tests against your current site
  • Performance monitoring
  • Recommendations
  • A/B tested

Best forOnline shops, media and learning sites.

Typical timeline: 4–8 weeks

Model training

We train a smaller model on your tone, format or task. It often matches a larger model for less.

What’s included

  • Example data prepared and cleaned
  • Training and comparison with the original
  • Test report
  • Set-up and version control
  • Training
  • Open models
  • Lower cost

Best forHigh-volume tasks where cost, speed or a set format matter.

Typical timeline: 3–6 weeks to set up, then monthly

Testing and monitoring

We keep AI features accurate after launch, with ongoing tests and alerts when quality, speed or cost slips. Every model version is recorded.

What’s included

  • Tests for accuracy, relevance and safety
  • Live usage monitoring
  • Instruction and model log
  • Cost and speed dashboards
  • Testing
  • Quality alerts
  • Version records

Best forTeams running live AI without quality tracking.

Your enquiry

Tick “Add to enquiry” on any offer, choose a plan, then continue. Or enquire about one offer.

How we work with you

Ways to start, from a question to a quote.

Ask a quick question, send your project details or request a quote. The person who will lead the work reads each one in full. Services in your enquiry go with it.

Or book a 30-minute call

What are you sending?

  1. 01

    About 2 minutes3 short answers

    For a first chat, a press request or anything not ready to plan yet.

    You get A reply from a named lead

  2. 02Recommended

    About 8 minutes5 short steps

    Goals, who it is for, budget and timing, so our first reply can outline the work.

    You get Options and prices after one call

  3. 03

    About 15 minutesYour documents attached

    Your documents, deadlines and the purchasing and security rules to meet.

    You get Confirmation and a named lead

How it is priced

Each plan shows how it is priced. Work starts once we agree a written scope and price.

  • Typical length
    1–3 weeks
    Pricing
    Fixed fee
  • Typical length
    4–12 weeks
    Pricing
    Fixed price
  • Typical length
    3–9 months
    Pricing
    Fixed price per stage
  • Typical length
    Ongoing · 6-month minimum
    Pricing
    Monthly fee
Compare what each plan includes
What each plan includes and who it suits
PlanWhat you getBest for
SprintA short, fixed job that answers one clear question.
  • Scope and result agreed before we start
  • A senior lead and the right specialists
  • A weekly review with you
  • A clear answer or a test version
An audit, a brand check or a test version of an app
ProjectOne clear job for one fixed price.
  • A written scope: what you get and how we agree it is done
  • A named project lead and a fixed team
  • A shared plan with checkpoint dates
  • Source files, yours once paid for
A brand identity, a website, a simple app or a set of tools
MilestoneA bigger build in stages you approve and pay for one at a time.
  • Stages, each with its own scope and approval
  • A go or stop review at each stage
  • Re-plan between stages as you learn
  • Pay for each stage you approve
Work too big for one contract, such as a large app
RetainerA monthly team that runs and improves what we built.
  • Team time set aside for you every month
  • Agreed response times for requests and fixes
  • A monthly review and a to-do list
  • Planned improvements, not only fixes
Monthly social media and ads, or care for live sites and apps

04.12Questions/faq

AI assistant and automation questions, answered.

Each answer links to the section of this page where you can see it working.

Not answered here?

Ask it on a call, and we will answer using your own documents and data.

Book a call

Any AI can, so we plan for it. Every sentence must point to a source, or a check removes it. If no source fits, the assistant says so and hands over to a person.

Sources on this page §04.2Answer checker §04.7Quality checks

Yes, and it shows which document each answer came from. Before it answers, it finds the right passages in your files. Use it when answers must follow your own policies or data.

Sources on this page §04.2Answer checker §04.6Steps

Yes. Each document keeps its access rights, so the AI never uses a passage the person asking could not open. Data stays in your cloud, or with providers whose terms forbid training on it.

Sources on this page §04.3Conversation §04.6Steps

The one that tests best on your examples for quality, speed and cost. Often a small model sorts questions and a larger one answers. Swapping is a settings change and a test run.

Sources on this page §04.6Steps §04.8Cost & energy

Yes. Models can run in Indian cloud regions, an Indian data centre or your private cloud. Before we build, we check the design against India’s data law (DPDP Act 2023) and RBI payment-data rules.

Sources on this page §04.6Steps

Use your documents when facts change and answers need sources. Train a model for style, format or one narrow task. Many setups use both.

Sources on this page §04.6Steps §04.8Cost & energy

We test it on examples built with your experts. Each answer is scored for sticking to its sources, relevance, safety and finding the right passages. People check the automatic scores. After launch, we score a sample of live answers every day, next to the business numbers the AI should move.

Sources on this page §04.2Answer checker §04.7Quality checks §04.11Outcomes

It must pass every check first. Then it gets 5% of traffic, and more only while answers stick to sources. If a number slips, the last good version returns automatically.

Sources on this page §04.7Quality checks §04.11Outcomes

It depends on volume, the model and the text per request. We estimate it in the test version, then cut it before launch with saved answers and small models.

Sources on this page §04.8Cost & energy

Yes. Open models can run in your cloud or on your own servers, using tools such as vLLM. Document search can run in PostgreSQL with pgvector, or a separate search database.

Sources on this page §04.6Steps

Tell us what you need built.

You will talk to the lead who would run your work.

Book a call

Three ways to start

We always agree a written scope and price before work begins.

Choose one of the three ways above