What three AI hackathons taught me about shipping instead of demoing
Between the Google Cloud Rapid Agent Hackathon, the Redrob Data & AI Challenge, and a couple of smaller builds, a few lessons kept repeating — mostly about what breaks the moment judges start clicking around instead of watching a script run.
A demo script is not a product
The failure mode I saw most in early hackathon builds — mine included — was optimizing for a rehearsed happy path instead of the moment a judge deviates from it. Saarthi AI's test suite (121 tests, plus a dedicated MongoDB MCP smoke test) existed specifically because five live third-party APIs during a judging slot is a lot of surface area for something to go sideways in front of an audience, and "it worked in rehearsal" isn't a standard worth trusting.
Determinism is a judging-round feature
Both Saarthi AI's risk score and RecruitRank AI's candidate ranking are built as deterministic, weighted, inspectable pipelines with an LLM layered on top for narration — not an LLM asked to produce the core number directly. That's partly about correctness, but just as much about being able to answer a judge's "why did this rank higher" question with a specific weighted term instead of a shrug.
Small teams need explicit ownership, not just enthusiasm
On Saarthi AI, having Saksham Pathak (Parthmax) own UI/UX and team lead, Urmila Saini own the agent tool orchestration and MongoDB MCP layer, Aishrica Dhiman own data analysis, and me own grounding and test coverage meant nobody was duplicating work in the 36 hours we actually had. The projects that stalled in earlier, less structured team efforts were the ones where two people quietly built the same thing without realizing it.
Ship the boring parts first
In every one of these builds, the unglamorous infrastructure — caching, fallback providers, OCR worker isolation, a real test suite — ended up being what made the demo survivable, not the flashiest model call. That's the pattern I'd tell anyone starting their first hackathon: the LLM call is rarely what breaks at 2am the night before judging; everything around it is.