Multi-agent clinical-trial integrity monitor on GCP

The scenario

Thousands of completed trials never publish results, and no human can read every registry record. TrialWatch ingests public registry and literature data into Postgres with pgvector, then runs specialist agents in parallel under a supervisor — each hunting one kind of integrity problem. Findings feed a memory system, low-confidence signals go to a reviewer, and the reviewer’s corrections improve the next run. Served as a FastAPI app on Cloud Run with infrastructure provisioned from the command line.

System architecture

High-level flow of TrialWatch, from input to output.

What you will learn

• Parallel specialist agents under a supervisor
• Episodic, procedural and semantic memory for agents
• Human-in-the-loop corrections that change future behaviour
• Provisioning and tearing down GCP resources cleanly

What you will build

• Registry and literature ingestion
• Six detection agents and a supervisor
• Reviewer queue with feedback into memory
• Cloud Run API with IAM and secrets

Build plan

1. Ingest registry data
2. Build two agents end-to-end
3. Scale to six with a supervisor
4. Add memory and the reviewer loop
5. Deploy and tear down to zero cost

Tech stack