Devpost submission draft — FinePrint

Track: Taskmaster (multi-step workflow automation) Repo: https://github.com/xiehuanyi/fineprint-agent Hosted URL: https://fineprint-ntemaj5ysa-uc.a.run.app/dev-ui/?app=fineprint Video: https://youtu.be/d0Mk5i7HpaY

Inspiration

This week I vetted 17 live hackathons by hand before entering this one. Three had prize pools that evaporated on reading the rules — a "$149,525" event with zero cash, a "$35,785" event whose top cash prize was $100, and a "$200,000 hackathon" that was actually a live-trading contest risking my own money. One "global" event required Singapore residency in paragraph nine. Reading fine print across three platforms took me a full day. That day is exactly what an agent should do instead.

What it does

FinePrint scouts open hackathons across Devpost, DoraHacks and TAIKAI, reads each event's actual rules page, and answers what the tiles never say: whether YOU can enter (country, student status, team size, payout rails), how much of the headline prize is REAL cash (credit inflation is endemic), whether it's even a build-and-judge hackathon, and which events deserve your calendar — as a tiered portfolio with deadlines and expected value per hour.

How we built it

Google ADK multi-agent tree on Gemini 3.7 Flash: a coordinator delegates to three specialists (scout / vetter / strategist) via description-driven transfer. All fetching is deterministic Python function tools — Devpost's JSON API, TAIKAI's GraphQL, DoraHacks through a rendering proxy (the site sits behind AWS WAF) — with 30-minute disk caching and fail-soft stale fallbacks, so a dead platform degrades the answer instead of killing it. The vetter is instructed to judge only from fetched text and to show its prize arithmetic. Deployed on Cloud Run via adk deploy cloud_run.

Challenges

DoraHacks' WAF blocks curl AND headless Chrome — solved with a rendering proxy plus retry/cooldown/stale-cache. Prize forensics needed careful prompting: the difference between "$10,000 in credits" and "$10,000" is invisible unless you force the agent to itemize cash line by line.

Accomplishments

The demo finds real events, catches real fake pools (the $149k → $0 case is reproducible live), and the whole build — three platform integrations, three agents, deploy — was a solo sprint inside the submission window.

What we learned

Agent judgment is only as good as tool honesty: returning warnings and stale-flags from tools, and telling the model to admit gaps, mattered more than any prompt cleverness.

What's next

Grant calls, CFPs and bug-bounty programs have the same fine-print problem; the vetter's kill-checks generalize. Watch-mode (daily diff + "NEW event you qualify for" alerts) is a natural Cloud Scheduler add-on.

Built with

google-adk · gemini-3.7-flash · Google Cloud Run · Python · Devpost API · TAIKAI GraphQL · r.jina.ai

Disclosures

Built during the submission period using AI coding assistants (permitted). The three platform fetchers were written earlier the same day (2026-08-30, in-period) as standalone scouting scripts and adapted into ADK tools.

Built With

• cloud-run
• gemini
• google-adk
• google-cloud
• graphql
• python
• vertex-ai