Deployed Clinical Trial Intelligence using GCP and LangGraph

MOSAIC is a production multi-agent clinical trial intelligence system built on GCP. It ingests real records from ClinicalTrials.gov and PubMed into Cloud SQL PostgreSQL with pgvector, then uses LangGraph to run six specialist agents in parallel under a supervisor — Broken Promises, Missing Results, Track Record, Pattern Finder, Side Effect Checker, Timeline Analyst — each hunting a different research-integrity problem. Findings go into a three-layer memory system (episodic, procedural, semantic), and low-confidence signals hit a human-in-the-loop gate whose corrections write back into procedural memory, so agents improve on the next run. The whole thing is served through a nine-endpoint FastAPI app, containerised, and deployed on Cloud Run, with Cloud Storage, Artifact Registry, Secret Manager, IAM service accounts and gcloud-driven provisioning behind it. It solves a genuine problem — ~30% of completed trials never publish results, and no human can read 400,000 records at once — and it produced real signals from live government data before the project was torn down to zero billing.