Design an on-prem agent for pharma deviation investigations

Anvika Pharma (fictional) makes generic oral solids and injectables at three plants in Hyderabad and one in Cork, Ireland. It supplies the US and EU markets, so its systems are inspected by the US FDA and EU regulators.

When something goes wrong in manufacturing (a tablet hardness result out of specification, a temperature excursion in a cold room), QA opens a deviation. An investigator then pulls data from several systems, works out the root cause, and writes a report with corrective and preventive actions (CAPA). Average closure time is 28 days; the target is 15. About 1,900 deviations a year.

The client wants an agent that drafts the investigation: gathers the relevant data, proposes likely root causes with evidence, and drafts the report for a human investigator.

Constraints:

• On-prem only. No cloud model calls. Available hardware: a cluster with 8 NVIDIA H100 GPUs in the Hyderabad data centre; Cork connects over a private WAN.
• Source systems: MES (batch records), LIMS (lab results), a historian for equipment sensor data, the QMS (past deviations and CAPAs), and SOPs in a document management system. The agent may read them; it must never write to them.
• Computerised systems must meet 21 CFR Part 11 / EU GMP Annex 11 expectations: every AI contribution must be attributable, time-stamped, reproducible and reviewable. The QA investigator signs the final report.
• Investigators have been burned by "confident nonsense" from an earlier chatbot. Credibility is fragile.
• The system must be validated before use in production (computer system validation).

Design the agent and how you would get it into production.