Consilium

The chart proposes, the room disposes.

Agentic treatment-decision support for relapsed/refractory aggressive B-cell lymphoma (DLBCL / transformed FL), second line, one decision point. Built for the Abridge × Anthropic × Lightspeed hackathon ("agentic AI in healthcare").

Every other oncology tool goes chart → recommendation. Consilium goes chart → room → recommendation. Two behaviors fall out:

1. A signal from the visit conversation can correct the chart — the record says one thing, the room says another, the plan changes.
2. When the clinician goes off-guideline defensibly, the agent names the boundary crossed, grounds the tradeoff in evidence, and lets them proceed. It explains; it never hard-stops a licensed physician. Hard stops are reserved for the affirmatively unsafe.

The screen is a temporal narrative — state changing over time as the room speaks — not a dashboard.

Pipeline

Three LLM agents plus two deterministic assets:

chart.json ─▶ chart.ts (deterministic, no LLM) ─┐
                                                ├▶ reasoner (LLM ×2: chart-only, then chart+room)
transcript ─▶ signals (LLM, verbatim spans) ────┘        │
                                                         ▼
                                     RecommendationSet {pre, post, delta}
                                                         │
                                     verifier (rule-check + in-context grounding) ─▶ VerifierReport

Retrieval lives only in the verifier (in-context over the evidence pack — no vector store). Cached (Cases A/B) and live (Case C) run the identical runPipeline.

Repository layout

src/
  app/            Next.js app router (single page + /api/run)
  lib/            contracts (zod), chart normalizer, rules, pipeline, anthropic client
    agents/       signals · reasoner · verifier
  components/     UI (temporal-narrative stage)
  types/
data/
  cases/          the two synthetic twin cases (+ jsonl)
  knowledge/      rules.json · evidence-pack.json · schema/  (runtime clinical brain)
  fixtures/       baked pipeline outputs (cached A/B)
docs/
  clinical/       rules.rubric.md · coverage.md · cases/   (Shalin's clinical source)
  evidence/cards/ per-source bibliographic cards
  design/         build-plan.md
scripts/          validate-knowledge · bake-fixtures
tests/
archive/          unused/superseded (25-encounter raw set, old evidence manifests)

Getting started

git clone https://github.com/VincentMao/HackathonProj.git
cd HackathonProj
npm install
cp .env.example .env.local   # add ANTHROPIC_API_KEY
npm run dev

Other scripts: npm run typecheck · npm run test · npm run validate:knowledge · npm run bake:fixtures.

Data & safety

All patient data is synthetic and de-identified — no real PHI anywhere. This repo is public: it never contains source PDFs or NCCN-derived text (license forbids reproduction); only the derived evidence-pack structure with citations and links. See .gitignore.

Team

• @VincentMao — engineering
• Shalin — clinical (Yale lymphoma oncology)