terryclinical-trial intelligence · headless

Competitive-landscape research in life sciences is slow, manual, and mostly reading. terry does the reading. You give it a condition and an intervention; it pulls the relevant trials and the literature around them, then hands back a single structured brief — not a pile of links.

It runs on its own. No dashboard to babysit, no human stitching sources together. Public data in, decision-support out.

What it does

• Searches ClinicalTrials.gov (API v2) for trials matching condition, intervention, and phase
• Enriches against PubMed (E-utilities) for the surrounding literature
• Synthesizes a structured JSON brief: findings, citations, confidence scores, and the data gaps it couldn't close
• Exports clean markdown, or delivers straight to Notion

How it's built

• Connectors — typed clients over the two public APIs, with a shared rate-limited HTTP layer
• Agents — an orchestrator drives enrichment → analysis → synthesis as discrete steps
• BYOK synthesis — pluggable LLM (mock / OpenAI / Anthropic); the default needs no key and runs rule-based, so it works offline
• Async by design — submit a brief, get a poll URL, collect the result; built to run unattended
• FastAPI · Python · Docker — one container, one command to run

Why it's shaped this way

Confidence and gaps are first-class, not decoration. A brief that quietly omits what it doesn't know is worse than useless in a clinical context — so terry is built to say "here's what I found, here's how sure I am, here's what's missing." Everything runs on public sources and stays local; nothing about a query has to leave the machine.

It was built in a single night — the week I decided to stop renting my tools and start owning them.

Status

Working prototype — live and public. Try the live demo or read the source.