Target safety assessment app (life-sciences pharma client)
biomedical-ai
knowledge-graphs
llm
rag
professional-services
Knowledge graph + LLM-driven evidence retrieval for regulator-ready target safety assessment reports. Anonymised.
What it is
A target safety assessment application built for a life-sciences pharma client under Biorelate’s Professional Services. The brief: take a candidate target and produce a regulator-ready safety report, faster and more reliably than the existing manual process.
The problem
Target safety assessment is evidence-heavy. Reviewers comb literature, clinical trial data, and curated databases to build a picture of a target’s safety profile. The manual workflow is slow, brittle, and hard to audit. Citations get attached after the fact, sometimes incorrectly.
What I did
- Led delivery end-to-end: scoping with the client, technical design, data integration, build, and handover.
- Wired Biorelate’s biomedical knowledge graph and curated literature corpus into an LLM-driven evidence retrieval layer.
- Made citation programmatic and deterministic, not LLM-generated. Every cited reference is provably present in the report’s reference list.
- Built non-Latin character sanitisation and citation-pattern validation into the export step. LLM hallucination patterns get caught before output reaches the client.
- Stood up a QA loop with the client’s domain experts to validate report quality against their evaluation rubric.
Outcome
- Regulator-ready safety reports, generated in a fraction of the manual review time.
- Citation integrity guaranteed at the system level (not LLM-dependent).
- Repeatable architecture: the same backbone now supports related assessment workflows.
Stack
Python, biomedical knowledge graph (Galactic), LLM evidence retrieval, programmatic citation, structured output validation, regulator-ready PDF export.