Inspiration
Emergency departments are one of the highest-stakes environments in healthcare. Clinicians have to make split-second decisions from fragmented information — vitals scattered across one screen, labs on another, imaging reports buried in a PDF, medication histories in yet another system. We wanted to build something that pulls all of that together instantly, without asking doctors to trust a black box. In high-trust clinical workflows, "the AI said so" isn't good enough — every recommendation needs to be explainable and auditable.
What it does
MedOS is an AI-powered emergency department operations platform that acts as an intelligent command center for clinicians. A user selects a synthetic high-risk patient, and with one click, five specialized AI agents run concurrently:
- Triage Agent — reviews symptoms, vitals, and risk indicators to propose an emergency priority
- Medication Safety Agent — detects allergy conflicts, drug interactions, and contraindications
- Lab Analysis Agent — summarizes abnormal labs, trends, and urgent results
- Imaging Review Agent — extracts key radiology findings and connects them to the patient's presentation
- Care Coordination Agent — synthesizes all outputs into a prioritized action plan (urgent / next / monitor)
Every recommendation is backed by evidence references from the patient record, and the full evaluation trail — confidence, latency, model version, evaluation status — is visible through Braintrust.
How we built it
We built MedOS with Next.js (App Router), TypeScript, Tailwind CSS, and shadcn/ui for the frontend, with Zod for structured, type-safe agent contracts. Server-Sent Events stream live agent progress to the UI so users can watch the five agents work in real time.
Fireworks AI powers every single agent invocation, giving us fast, reliable inference across all five agents running concurrently. Braintrust is the core differentiator of the product — every agent run creates one root trace with five child spans, each carrying deterministic safety evaluations and operational metrics. This isn't a model self-reporting confidence; the confidence score is derived from actual evaluation results, and any recommendation that fails a hard safety check is automatically blocked or flagged for human review.
All patient data used in the demo is synthetic, and we used static fixtures rather than a live EHR connection to keep the hackathon build focused and safe.
Challenges we ran into
Coordinating five AI agents to run concurrently while keeping their outputs consistent and mergeable into a single action plan was the biggest technical challenge — we needed structured, predictable contracts (via Zod) so Care Coordination could reliably consume four different agents' outputs. Building genuine trust and observability, rather than just displaying a confidence number, also took real design work — we wanted evaluation, not vibes.
Accomplishments that we're proud of
We're proud that MedOS never just "trusts the model." Every recommendation is traceable, evidence-backed, and blockable if it fails a safety check. Getting five agents to process concurrently and stream live progress to the UI, all while wiring in a full Braintrust observability layer, in the hackathon timeframe, felt like a real win.
What we learned
We learned how much harder — and more important — the trust layer is compared to the AI layer itself. It's relatively straightforward to get a model to generate a recommendation; it's much harder to prove that recommendation is safe, explainable, and auditable enough for a clinician to actually act on it.
What's next for MedOS
Next, we want to connect MedOS to real EHR systems via FHIR APIs, add authentication and role-based access for real clinical environments, and expand the agent roster with things like allergy cross-referencing and discharge planning. Longer term, we're interested in clinical validation studies and exploring pilot programs with real health systems.
Built With
- braintrust
- copilot
- firework
- next.js
