MedBot Hack the North · 2026
Product engineeringApplied AI · UX design
Turning short-form health claims into an evidence review people can actually follow.
View MedBot on DevpostFour people. 36 hours.
1st place, GPTZero track.
A confident interface can make an uncertain answer feel definitive. The design needed to help people examine the evidence, not simply accept a verdict.
The work behind the interface
I formed and coordinated our four-person team, led product and UX direction, and built React/TypeScript product flows. I also contributed to Python/FastAPI pipeline integration and front-end/backend data contracts. The full-stack system was a shared team build; my work connected the product experience with the evidence pipeline.
Set supported-source and clip limits before the request starts.
Show progress, extracted claims and recoverable failures.
Keep source passages, applicability and limitations beside the assessment.
Written findings remain useful while the longer media step continues.
Product journey from the team implementation. This diagram describes the flow; it is not a set of captured result screens.
My product, front-end and integration work connects to a shared evidence pipeline. The current team repository implements the flow below; these are code-level capabilities, not a claim that every service was deployed at the hackathon.
GPTZero checks authorship separately from medical judgments. Valid citations do not guarantee a correct interpretation. That boundary matters in both the interface and the implementation.
Architecture verified against the team repository, 23 September 2026. Live service configuration and medical-verdict accuracy are not established by this code review.
Hack the North submission. Example outputs, not an accuracy benchmark. Watch on YouTube ↗
Written findings appear before the generated explainer is ready. The result is useful while the longer media step continues.
A video is another way into the information, not a gate in front of it.
The experience moves from the original claim to an understandable finding, with source exploration and limitations close by. Progress states make the long-running analysis visible.
The reader should be able to inspect why a conclusion was reached and where it may fall short.
Whether the transcript shows signs of AI-generated writing is a different question from whether its health claim is supported. These signals stay separate in the product framing.
A real person can repeat a false claim; synthetic media can contain a supported one.
Outcome & limits
We shipped the prototype in 36 hours and won 1st place in the GPTZero sponsor track at Hack the North 2026. My front-end contribution was merged into the team repository. The mobile interface gives the input a clear starting point and keeps written findings useful alongside the captioned explainer concept.
Medical-verdict accuracy was not formally validated, and some judgments were incorrect. This is a hackathon research prototype, not a clinically validated tool or a substitute for professional advice.
Shared full-stack prototype; I led product/UX and contributed front-end and integration work
Hack the North 2026; scope, not a performance measurement
GPTZero track—not an overall hackathon win or accuracy benchmark
Proposed · What I’d explore next
Let an unresolved evidence gap trigger one more targeted lookup, then return a reviewable draft with a clear stopping point. This is a proposal, not an implemented agent.
Compare the current baseline, optional reranking and the agent on frozen, human-labelled failure cases. Track citation correctness and coverage; applicability and abstention; p50/p95 latency and cost.
Build on the existing FastAPI / Celery pipeline, model adapter and Pydantic schemas with named tools and persisted action events. Preserve unstated claim fields as unknown. Inspect the cited study and independent Europe PMC evidence; record abstract-only versus full-text access. Use Elasticsearch’s existing BM25 + ELSER / RRF query. Let evidence gaps drive additional lookups, including contradictory findings, within tool, time and cost budgets. Deduplicate sources, retain correction metadata and validate stable passage IDs and exact quotes. Quote validation cannot establish medical correctness. Save a versioned draft and an action log for human approval, not private model reasoning. Optional reranking already exists and defaults off. Consider LangGraph only if durable branching becomes necessary; Agent Builder and Elastic Workflows are not confirmed deployments, and MCP would be a tool interface, not the retrieval method.