Why I built it
After the OpenAI x START Warsaw hackathon, I had applications, attendance records, project submissions, GitHub links, and review notes. The basic numbers were easy to count. The harder question was: what did this community actually build, and what could I show a partner considering the next event?
Most communities answer that manually with spreadsheets. I wanted a system that could produce useful, repeatable evidence instead of another one-off report.
What I built
Community OS imports supported Luma and Devpost exports, reconciles people across those sources, and sends ambiguous matches for review.
Optional enrichment adds bounded public project evidence. Before GPT-5.6 sees anything, the pipeline removes direct identifiers, contact details, profile links, secrets, and repository paths. GPT-5.6 proposes structured classifications, but a human reviews them. Deterministic code, not the model, calculates the final metrics.
The approved data produces an interactive partner dashboard and a fixed PDF. For another supported hackathon, the organiser changes the event details and selects new exports. The same adapters, review gates, metric definitions, and renderers can be reused without event-specific model fine-tuning.
Live result
The published report covers 286 applicants, 83 accepted participants, and 78 confirmed attendees. Partners can switch between those cohorts and inspect what people built, their technical and product evidence, founder experience, customer delivery, and combinations of those signals.
Every metric keeps its definition and denominator. Missing evidence remains unknown rather than becoming a negative score.
How I used Codex and GPT-5.6
I started an initial ingestion prototype on July 11, before Build Week officially opened. During Build Week, I substantially extended it into the reusable reviewed pipeline, private operator, responsive dashboard, PDF, synthetic testing path, publication controls, and GitHub-to-Vercel release.
Codex with GPT-5.6-sol was my main engineering environment. I used it to turn product decisions into failing tests, implement the pipeline and interface, recheck the headline counts independently, inspect desktop and mobile layouts, review privacy boundaries, and verify the production release.
GPT-5.6 also runs inside the product, but in a limited role. It proposes structured classifications from approved evidence. It does not rank participants, calculate the final statistics, or publish the report.
What I learned
The difficult part was not creating charts. It was keeping every number tied to a reviewed source, a clear definition, and the correct denominator.
One production check also found that analytics was still deriving location fields. I stopped the release, disabled that enrichment explicitly, added a regression test, and verified fresh events before publishing.
The next test is simple: run the same pipeline on another compatible event and see how little needs to change.
Log in or sign up for Devpost to join the conversation.