Asia PacificAn expert's know-how, test results and development records, in one assistant that cites every answer
The division's know-how was spread across an external expert's lectures and slides, partner test results, technical feedback and internal development records, and staff needed answers they could repeat to customers. In a Craft Sprint we proved, on a first part of the lectures and slides, that a confidential assistant could answer with exact sources and decline when the material has no answer. It is now being built on Amazon Bedrock in a closed AWS environment that brings test results, feedback and development information into the same place.
kinds of know-how in one cited source: expert theory, test results, partner feedback, development records
Expert test questions answered correctly in the trial
- Client
- Engineering division of a Japanese industrial manufacturer
- Users
- Division staff: 5 in the trial, with wider access planned at production scale
- Scope
- Craft Sprint prototype · Trial · Production design on AWS
- Timeline
- Aug 2026 prototype → Aug 2026 trial → production build
- Content
- Trial: part of the lecture series (15 recordings) and 39 slide decks. Next: test results, technical feedback and development records
Answers with a source citation in the 22-question test set
Out-of-scope questions correctly declined rather than guessed
Specialist knowledge scattered across video, slides and test reports
The division sells a continuous purification system to biopharmaceutical manufacturers. Much of its know-how on antibody purification, chromatography and process scale-up came from a long lecture series given by an external bioprocess expert, recorded on video with accompanying slides and exercises. Just as important were the results of evaluations run with partners, technical feedback from outside experts and the division's own development records, each kept in a different place.
Staff needed answers they could repeat to customers. A general-purpose AI chatbot could not help: it does not know the unpublished material, it cannot show where an answer comes from, and it fills gaps with plausible guesses.
Knowledge in too many places
Lectures, slides, test results, feedback and development notes sat in different places, and the recordings could not be searched at all.
Answers that must be traceable
Anything said to a customer had to point back to a specific lecture or slide.
Confidential material
The recordings and slides could not leave the company's control or be used to train a public model.
Prove trust on real material, then bring every source of know-how into one place
We built an anonymous assistant, not a digital copy of the expert. It answers only from the indexed material, cites every sentence and replies "not in the material" when it has no basis. Recordings are transcribed and indexed on a local machine; only short text excerpts are sent to the language model, under commercial terms that exclude training.
Craft Sprint prototype
- Local transcription with Whisper, 4.6× faster than real time
- Glossary and slide-aware correction of misheard technical terms
- Hybrid keyword (BM25) and semantic search with content-hash incremental ingestion
- 22-question test set, rerun after every content addition
Trial
- Web app for a 5-person trial behind email-based access control
- Usage log of every question, to find gaps in the material
- Optional web mode limited to allow-listed regulatory and scientific domains, with internal and public sources labelled separately
Operating options
- Briefing deck for decision makers: how the assistant differs from general AI, live demo, three hosting options
- Measured cost per question for the fast and detailed answer modes
Production design
- AWS Tokyo Region in a dedicated account, reachable only over VPN from the client network
- Claude via Amazon Bedrock over PrivateLink; containers across two Availability Zones
- Automatic ingestion for new documents, test results and development records
- 8–10 week plan including parallel running and a quality check against the trial
From scattered material to one source engineers and sales can rely on
The most important answer this assistant gives is "that is not in the material". People will only trust it in front of a customer if it never invents a fact.
From a lecture library to the division's engineering knowledge base
The trial ran on a single machine with part of the lecture material. The production build moves the assistant to Amazon Bedrock in a dedicated AWS account in the Tokyo Region, closed to the internet and reachable only from the client's network. Through the same entry point it will add test results with their conditions, technical feedback from partners and internal development information, so that engineers can prepare tests and answer customers from one source, with theory and measured results kept clearly apart.
Closed environment
Dedicated account, VPC and site-to-site VPN, single sign-on, encryption and audit logging defined as code.
Rebuild and verify
Re-index with Amazon Transcribe and Bedrock embeddings, then compare answer quality on the same questions as the trial.
Scale the content
Add the remaining lectures, test results, partner feedback and development records, and extend access to the wider division.
How we can help with a similar problem
Facing a similar challenge?
Tell us about it in a few lines. We will come back with a practical first step, usually within a week.