Customer background
A multinational generics manufacturer with multiple R&D sites across Asia-Pacific, holding over 20 years of experimental records, clinical trial data, regulatory filings and process-optimization reports.
This knowledge was scattered across: email attachments, an internal SharePoint, a legacy LIMS, scanned paper records, and several senior researchers' personal machines.
The problem: knowledge locked in silos, new hires starting from scratch
Every time a senior researcher left or retired, their replacement typically needed 6–12 months to "get up to speed." Even for existing staff, common friction included:
- Knowing a particular experiment was done three years ago, but being unable to find the report
- Needing historical stability data for a compound, but not knowing which system it lived in
- Preparing regulatory filing materials and having to re-assemble previously assembled knowledge every time
What we built
Using the LowsiteOS RAG (Retrieval-Augmented Generation) knowledge layer as the core, the project built a private, on-premises-only knowledge base queryable through natural language.
Phase 1 — Document inventory and cleaning
Mapped the 20-year document library: identified key sources, established document classification and versioning strategy, and configured access controls per confidentiality level. This phase was led by the customer's IT and regulatory teams; MICH provided the tooling and process framework.
Phase 2 — Vector index construction
All documents were OCR'd, structurally chunked and vector-embedded entirely on-prem via the LowsiteOS document parser. No network connection required, no third-party API called at any stage.
Phase 3 — Query interface launch
Researchers submit natural-language questions through a web interface. The system returns integrated answers with source document links and page numbers for every cited passage, ensuring every answer can be traced back and verified.
