Clinical Trial Band
Clinical Trials is a multi-agent system for Track 3 (Regulated & High-Stakes Workflows) that matches metastatic breast cancer patients to recruiting clinical trials. Four specialized AI agents coordinate through Band as a single team: -Intake (Featherless · Qwen2.5-32B-Instruct) normalizes the patient profile — receptor status, ECOG, prior therapies, labs. -Discovery (Featherless · Qwen2.5-32B-Instruct) scans curated ClinicalTrials.gov data for candidate trials. -Eligibility (Featherless · Qwen2.5-32B-Instruct) restates each trial's free-text criteria into structured constraints. -Analysis (AIMLAPI · GPT-4o-mini) writes a per-criterion clinical rationale and publishes the ranked result. Agents hand off through @mentions in a shared Band room; the dashboard live-tails the conversation so a clinician watches the agents collaborate in real time. The core architectural principle: the LLMs never compute the verdict. Match / no-match for every criterion is calculated in Python from structured trial levers. The models explain in clinical language — "HER2-positive matches the trial's HER2-expressing requirement" — but they can never change a pass/fail. Same patient in, same verdict out, every run. That's what makes agentic AI safe enough to deploy in a regulated workflow. Why it's original: Agents on two different model backends (Featherless + AIMLAPI) coordinate as one team through Band. The Python decision engine sits as the architectural boundary the models never cross — a guardrail by design, not a post-hoc reviewer. The same human-in-the-loop, explainable pattern that leading cancer centers deploy at scale. The stack. Band (coordination · @mention handoffs · task-state lifecycle · retry) + Featherless (open-weight reasoning agents) + AIMLAPI (analysis) + LangGraph + FastAPI + React/TypeScript. Real ClinicalTrials.gov data, synthetic patients, no PHI.
Category tags:.png&w=96&q=75){kind=small}
"Application of technology: Four metastatic breast cancer clinical trial matching agents (Intake/profile normalization, Discovery/ClinicalTrials.gov scanning, Eligibility/free-text criteria structuring, Analysis/per-criterion GPT-4o-mini rationale) on two model backends (Featherless Qwen2.5-32B + AIMLAPI GPT-4o-mini). Python decision engine computes match/no-match from structured trial levers — LLMs explain but never compute the verdict. Same input → same verdict every run. Real ClinicalTrials.gov data. Presentation: Precise architectural principle: "The LLMs never compute the verdict — match/no-match for every criterion is calculated in Python from structured trial levers." "The models explain in clinical language but they can never change a pass/fail." This is deterministic reproducibility applied to regulated healthcare AI. Solo submission. Business value: Clinical trial matching for metastatic breast cancer patients is life-critical. Python verdict engine ensuring same-input/same-output determinism makes this safe enough for regulated deployment. Per-criterion clinical language rationale satisfies clinician explainability requirements. Originality: Python decision engine as an architectural boundary that models can never cross — a "guardrail by design, not a post-hoc reviewer" — is one of the clearest and most precise statements of safe AI architecture in the hackathon. Cross-backend coordination (Featherless open-weight + AIMLAPI proprietary) through Band for a single clinical task demonstrates Band's multi-provider value."
Sanem Avcil
For Innovators & Creators