Pharma and Life Sciences · AI
Private AI assistant for a pharma quality team to search SOPs, deviations and audit records
A private knowledge assistant that answers questions from approved SOPs, deviation reports, CAPAs and past audit responses, citing the exact document and version for every answer. It runs inside the company's own cloud, so no quality data leaves its control.
- Client
- A mid sized formulations manufacturer in Gujarat supplying regulated export markets
- Region
- India
- Sponsor
- Head of quality assurance and the site QA manager
- Duration
- 16 weeks to first release, including validation, then quarterly reviews
Engagement model
FDE: Forward Deployed Engineering
Two engineers worked on site with the QA team for the build, because document access, validation and change control had to follow the plant's own procedures.
How FDE worksThe team
- Engagement lead1
- AI engineer2
- Backend engineer1
- Validation and QA analyst1
The challenge
Where things stood
- The QA team held thousands of SOPs, deviation reports, CAPAs and audit responses across a document system and shared drives.
- Before each regulatory audit, staff spent days finding how a similar observation had been answered before.
- New QA staff struggled to find which SOP version applied to a task and what had changed.
- Leadership would not allow quality records to be sent to a public AI service.
The approach
How we ran it
Weeks 1 to 3
Scope and controls
Agreed with QA which document types were in scope, who could see what, and how the assistant itself would be validated and kept under change control.
Weeks 3 to 8
Index the records
Loaded approved and effective documents only, kept version and effective date with every passage, and applied the same access groups as the document system.
Weeks 7 to 12
Build the assistant
Ran an open weights model inside the client's Azure tenancy, built answers that always cite document number and version, and made the assistant say so when the records do not cover a question.
Weeks 12 to 16
Validate and release
Wrote test scripts with QA, ran them against a fixed question set, recorded the results as validation evidence, and released to the QA department first.
What it looks like
Representative screens
Drawn to show the kind of system we delivered. Client screens and data stay private.
QA assistant
What is the hold time for cleaned granulation equipment before re-cleaning is needed?
Per SOP QA-CL-014 version 6, effective 12 March 2025, cleaned equipment must be used within the clean hold time stated in Annexure 2 for that equipment, otherwise it is re-cleaned and the logbook updated. Annexure 2 lists the rapid mixer granulator RMG-02 separately.
Have we had an audit observation on this before?
Yes. Audit response AR-2023-07, item 4, answered an observation on clean hold time records for RMG-02. CAPA-23-118 added a logbook column for clean hold expiry.
Sources: QA-CL-014 v6, AR-2023-07, CAPA-23-118
What we delivered
Modules and features
- Private language model hosted inside the client's own cloud tenancy
- Search and question answering over SOPs, deviations, CAPAs, change controls and audit responses
- Citations with document number, version and effective date on every answer
- Access rules mirrored from the document management system
- Audit preparation view that lists past observations and responses by topic
- SOP version comparison that summarises what changed between versions
- Full log of questions, answers and sources for review
What changed
Outcomes
- QA staff find the governing SOP and its current version in a single question
- Audit preparation starts from a ready list of past observations and answers
- New team members learn procedures with the source document in front of them
- Quality records stay inside the company's own cloud environment
Looking back
The question that mattered most to QA was whether the assistant could ever quote a superseded SOP. Loading only effective documents and showing the version on every answer settled that, and we would make it the first design decision again. Validation took longer than the build, which was right for this setting, and we should have planned the test scripts with QA from week one. The QA team responded well to the assistant stating plainly when it had no source.
Start a conversation
Discuss a similar project
Tell us where your team is today and we will walk you through how a project like this would run for you under FDE. The first conversation is free. Delivery work is paid only after you accept it.
A specialist replies within one working day. The first conversation is free.