Coverage note · Applied AI — other domains · Healthcare · Agentic AI
CADUCEUS — Virtual Molecular Tumor Board
Hand it a full oncology case and 55+ specialist agents across 7 layers convene a molecular tumor board — diagnosis, therapy, pharmacy, trials, evidence, dissent — and return a fully-cited, guideline-concordant recommendation with a human on the final gate.
Key metrics
1. Hierarchical supervisor over departmental sub-graphs; the four diagnostic departments run as one parallel wave, while the Variant / Trial / Abstract agents fan out one worker per item (~150 concurrent at peak). A dedicated Dissent / Red-Team agent argues against the consensus and a Citation Validator blocks any unsourced claim — 100% citation grounding. Decision support, not decision making; built only on free, public, synthetic or de-identified data; not an FDA-cleared device.
The problem
A molecular tumor board is the multidisciplinary review every cancer centre runs weekly: pathology, radiology, genomics, labs and notes fused into one defensible plan. It is the textbook case for multi-agent AI — and the textbook case for getting AI in medicine wrong.
The three things that sink clinical AI are exactly what this is built around: no provenance (nothing is trustable), silent consensus (no one argued the other side), and automating a decision that must legally and ethically keep a human on the hook.
Architecture
- 01
Ingestion & medallion
Bronze raw -> Silver FHIR-normalised -> Gold unified Patient Case object the agents consume. Synthetic Synthea-style cases, or a real de-identified TCGA patient assembled live from the NCI GDC open tier.
- 02
Knowledge layer
Hybrid retrieval: a vector view over literature / drug labels / guideline text and a knowledge graph of gene -> drug -> disease -> interaction relationships, every chunk carrying a source id.
- 03
55+ agents, 7 layers (LangGraph-style)
Hierarchical supervisor (Board Chair) over departmental sub-graphs: Intake, Diagnostic (Pathology / Radiology / Molecular / Labs), Therapeutic + Pharmacy, Evidence / Trials / Guidelines, Supportive, and Synthesis.
- 04
Genuine concurrency
The four diagnostic departments run as one parallel wave; Variant Actionability, Patient-Trial Matcher and Abstract Screening fan out one worker per item (map-reduce) — peak concurrency ~150, streamed live to the UI over SSE.
- 05
Synthesis & safety
A Consensus Synthesizer fuses modalities; a Dissent / Red-Team agent argues against the recommendation; a Contradiction Checker and a Citation Validator block unsourced or conflicting claims before a human sees them.
- 06
Live data path
On real patients: variants annotated against CIViC, drug toxicity from openFDA + FAERS, trials from ClinicalTrials.gov v2, literature from Europe PMC — all keyless free public APIs, with honest open-tier data gaps marked, never fabricated.
Key tradeoffs
Provenance or it did not happen — a Citation Validator rejects any recommendation lacking a retrievable source.
WhyIn oncology a confident-but-wrong claim is a liability. No source, no statement; 100% of claims trace to a citation.
A dedicated Dissent / Red-Team agent exists solely to attack the consensus.
WhySilent agreement is the failure mode of committee AI. The minority view and modality conflicts are surfaced, not smoothed over.
Decision support, not decision making — human-in-the-loop checkpoints after diagnosis and before finalisation.
WhyClinical AI that automates the decision is what regulators and clinicians reject. CADUCEUS proposes; a clinician disposes.
Keyless and free — no LLM API key, no credentialed data; real patients come from the TCGA open tier.
WhyA pipeline anyone can run on real public data beats a demo on invented data behind a paywall. Built only on free / synthetic / de-identified sources.
Eval results
Every clinical claim carries a retrievable source id, verified by the Provenance / Citation Validator across full runs; unsourced claims are blocked from output.
Inter-department parallelism plus per-variant / per-trial / per-abstract map-reduce fan-out; measured at the heaviest screening window.
Synthetic cases scored against NCI PDQ for matched cancer types; on live TCGA cases this is shown honestly as evidence-grade strength, since PDQ has no machine-readable API.
GDC / TCGA, CIViC, openFDA + FAERS, ClinicalTrials.gov v2, Europe PMC and RxNorm — all reachable with no key, verified live.
Production proof
The artifact that keeps the numbers honest: the eval harness and monitoring gates that run in CI, not a one-off notebook result.
Provenance + dissent + human gates by construction
CI · PASSINGCADUCEUS drafts; clinicians decide. Every recommendation has been argued against and every line traces to a public source. A research prototype, not an FDA-cleared medical device.
Pre-assembles the case, the evidence and the trial matches a tumor board spends hours gathering — with the provenance, documented dissent and human gating a risk-conscious clinical leader actually requires — on free, public, real or synthetic data.
Request coverage
I am focused on finance AI: credit risk, RegTech, AML, and agentic investment research. Open to roles, mentorship, and collaborators in fintech, quant, and bank AI.