Consilium
Consilium is agentic treatment-decision support for relapsed/refractory aggressive B-cell lymphoma that goes chart to room to recommendation to action.
Core trick
A deterministic rule engine ranks lymphoma regimens, and a verifier grounds each rationale in cited literature or fails loudly.
Description
Consilium is agentic treatment-decision support for relapsed/refractory aggressive B-cell lymphoma that goes chart to room to recommendation to action. A signals agent extracts the intangibles from the ambient visit (with verbatim quotes); a reasoner ranks the committed regimen set through a deterministic clinical rule engine, weighing chart and room together and flagging defensible off-guideline choices; a verifier grounds every rationale in curated primary literature and surfaces per-plan attention flags, failing loud if it cannot verify. The clinician selects plans and a summarizer produces an operational to-do list (orders, consents, referrals, monitoring) to send to downstream workflow. A signal from the room can correct the chart.
Live Vercel link: https://hackathon-proj-phi.vercel.app/
README excerpt
# 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
`