01 / The brief · Hack the North 2026 · 4 people · 36 hours
MedBot
Turning an AI pipeline into a product people can question.
Health claims spread faster than their evidence. We built a path from a short video to research people could inspect.
Open a recorded check- My role
- Product design + full-stack integrationI formed the team, led UX and built the front end, then contributed Python retrieval and video fixes.
- What we delivered
- An end-to-end AI research prototype.1st place, GPTZero sponsor track.
My front-end + retrieval work
- React
- TypeScript
- Python
Shared API
- FastAPI
02 / From problem to product
How we chose the solution
I prioritized written findings, quotations and source limits over one confidence score. Users could inspect evidence while video rendered.
How we built it
I connected React to the team API, then contributed Python retrieval and video fixes. The shared pipeline extracts claims, retrieves research and assesses evidence; teammates built its core services.
Try thisChoose a claim, open its cited paper and compare it with the review. Does the evidence support the assessment?
Open the evidence.
Saved run · 20 September 2026Portfolio replay of the product flow. No live medical analysis.
From claim to source.
Open the check, choose a claim, inspect its papers.
This run contains reasoning errors. Their review stays beside each finding.
Watch the submission ↓Open the saved check, choose a claim, inspect its papers. The review flags this run’s reasoning errors.
My code + the shared architecture
AI orchestration: extract claims, retrieve research, assess the evidence, then optionally render an explainer. These are coordinated pipeline stages, not autonomous agents. Written findings remain available while video renders.
React → FastAPI → Celery / Redis → retrieval + model services → saved results in PostgreSQL
- My React integration: case API, reconnects and guards against stale updates.
- My Python changes: claim/citation contracts, applicability fields and retrieval safeguards.
- My video fix: preserve fractional scene durations with frame-level regression tests.
Teammates shared API, queue, storage, model and GPTZero services. I integrated the approved UI; I did not independently build every component or service. GPTZero authorship detection is separate from medical evidence.
Recorded PR checks passed frontend, backend and video jobs. A separate Workers build failed; this does not verify live deployment or medical accuracy.
Two interface decisions, shown in the real React UI
Watch the hackathon submission
Product walkthrough; not an accuracy evaluation. Watch on YouTube ↗
03 / Findings + next step
What we found
We won the GPTZero sponsor track with a working prototype. My takeaway: good AI UX makes uncertainty visible. The saved run’s reasoning errors show why evaluation comes next.
What I’d improve next
Test claim-to-source accuracy on a reviewed set of examples before expanding the experience.