40% Fewer Batch Deviations with GxP-Compliant AI

Client

Who We Worked With

A mid-sized Indian pharmaceutical formulations manufacturer with multiple WHO-GMP-certified facilities, producing solid oral dosage forms for regulated export markets including the US, EU, and Japan.

Problem

The Business Problem

Repeat batch deviations were driving up rework, scrap, and regulatory exposure. Quality investigators were spending the majority of their time searching across PDFs, paper batch records, SOPs, and historic CAPAs to identify root causes - work that delayed CAPA closure and frequently missed analogous past events.

Baseline

Where They Started

• Average of 18 to 22 batch deviations per month across the facility
• Mean CAPA investigation cycle of 21 days from initiation to closure
• Root-cause analysis dependent on tribal knowledge of senior QA staff
• No systematic linkage between current deviations and historic precedents

Data

Data Sources Used

• 10 years of historic batch manufacturing records (PDFs and scanned)
• LIMS test results and out-of-specification (OOS) records
• QMS deviation, change control, and CAPA records
• Site SOPs, master batch records, and validation reports
• Regulatory references including 21 CFR Part 11 and ICH Q7/Q9/Q10

Solution

What We Built

We built a retrieval-augmented assistant trained on the client's anonymised historic batch and quality data, deployed entirely inside their private cloud. Investigators query the system in natural language during a deviation; the assistant surfaces the most similar past events, the CAPAs that closed them, and the SOP clauses involved - with every answer linked to a source document and a confidence indicator. A human approval step is mandatory before any record is updated.

Integration

Systems We Connected

• Read-only connectors to LIMS and QMS via existing APIs
• Document ingestion pipeline for batch records and SOPs
• SSO with the client's identity provider for role-based access
• Immutable audit log written to a tamper-evident store
• No data leaves the client VPC - all model inference runs locally

Timeline

Project Timeline

Discovery and data scoping completed in 4 weeks. Pilot deployment on a single product line in 8 weeks. Full facility rollout, including IQ/OQ/PQ qualification and staff training, completed in 5 months from kickoff.

Governance

Security & Compliance

• 21 CFR Part 11 compliant audit trail for every assistant interaction
• Role-based access control aligned to QA, production, and regulatory roles
• Human-in-the-loop approval before any QMS record is written
• Quarterly model performance and bias review by client QA leadership
• Full data sovereignty - client owns the model weights and training data