Context
The project explores how an AI-assisted review workflow can help organize variant evidence without asking the user to blindly trust a generated answer.
Back to portfolio
Case study
OncoTriage AI: provenance-first variant review
A genomic variant review copilot designed around explicit safety boundaries, evidence visibility, and failure handling.
Problem
In a sensitive domain, the hard part is not only calling a model. The workflow needs to show what evidence was retrieved, what the model inferred, and what failed or fell back.
My role
I designed and implemented the Python review flow, Streamlit interface, evidence integrations, mode handling, and provenance-first presentation.
Constraints
External evidence can be missing or unavailable, model output is not clinical evidence, and the project is a research-oriented portfolio system rather than a deployed medical product.
Architecture
Evidence lookup and model scoring are separate paths. The UI presents provenance and mode state so users can distinguish retrieved evidence from generated interpretation.
Important decisions
I used explicit LIVE and DEMO modes, visible fallback behavior, and provenance-first output structure to avoid making the model appear more authoritative than the supporting data.
Tradeoffs
Streamlit kept the review workflow fast to build and inspect, but it limits frontend control. Separating retrieval and scoring added orchestration work while making failures and provenance easier to audit.
Failure modes handled
The system accounts for unavailable evidence, inference failure, and demo-mode execution by making those states visible instead of substituting silent placeholder answers.
Outcome
The project demonstrates my preferred pattern for applied AI: keep uncertainty visible, keep boundaries explicit, and design the interface around review rather than automation theater.
Current limitations
The system is not clinically validated or deployed for patient care. Evidence parsing and fallback paths need deeper automated coverage, and the portfolio still needs production-safe screenshots.
What I would improve next
The useful next step is hardening, not bigger claims.
- Add focused tests around evidence parsing and fallback states.
- Expand case-study screenshots once production-safe sample data is available.
- Tighten the review workflow with clearer confidence and source grouping.