AHMED IRFAN

A Clarity agent decides if a question is specific enough — and if it isn't, the graph pauses and asks before continuing. Research, Validator, and Synthesis agents route between each other on confidence scores, retrying until the evidence holds. Answers stream word-by-word over Server-Sent Events, with a live tracker showing which agent is active.

Ten authenticated n8n workflows form a true orchestration: an intent classifier fans a query out to pricing, FAQ, general, and booking agents in parallel, then merges them into one reply — with a real collecting-to-confirming booking state machine that carries slots across turns. Multi-tenant by design: point it at any clinic URL and it re-ingests end to end.

A two-stage pipeline: one model call extracts a structured candidate profile, a second scores it against the role on a weighted rubric with a written recommendation, not just a number. Gmail attachments are picked up automatically; a resilient batch never lets one bad resume block the rest. An AI email assistant drafts replies grounded in each candidate's report — but nothing sends until HR clicks Send.

Built for the National AI Hackathon '26. Agents run GRIM checks and statcheck p-value recomputation, verify DOIs against Crossref, score reproducibility and methodology against CONSORT/STROBE/PRISMA checklists, predict replication, and scan for conflicts of interest — every finding flagged for human expert review, never treated as a definitive judgement.