Winner pages reveal the constraint
Devpost's 2024 AI hackathon roundup lists PartyRock, Microsoft AI Classroom, and Google AI Hackathon as major examples. PartyRock had 7,660 participants, Microsoft AI Classroom had 3,706 participants, and Google AI Hackathon had 15,587 participants.
The winner pattern is not just 'use AI'. Parable Rhythm won PartyRock overall first place by making an interactive crime-thriller game. ChatEDU won the Microsoft AI Classroom Technical Project Prize by turning notes and files into study guides, quizzes, and tutoring sessions. Nested won Google AI Hackathon overall first place by using Google Maps history to recommend local places after a move.
Narrow jobs beat generic agents
Microsoft's AI Agents Hackathon says 18,000+ developers registered and 570 submissions came in. Its Best Overall winner, RiskWise, is not a generic assistant. It analyzes supply-chain risk from shipping, news, and data, then lets analysts ask targeted questions.
The same Microsoft showcase names WorkWizee, TARIFFED!, Konveyor, Apollo, and ModelProof as category winners. Each one has a job: incident-call automation, tariff analysis, knowledge transfer, research synthesis, or answer auditing. That is the agent pattern to copy.
Physical proof still travels
The TreeHacks 2025 gallery marks winners such as HawkWatch, BrailleBot, BlinkAI, Iris, SurviveX, PUSHPA, ECGo, Therms, and WingNote. Many of these projects make AI visible through cameras, wearables, drones, printers, or sensors instead of hiding everything behind chat.
That matters for a judge room. A physical or multimodal demo gives the audience a before state and an after state. HawkWatch turns camera feeds into VLM alerts. BlinkAI turns blinks into speech. BrailleBot turns text into braille output. The demo object carries the story.
Big fields still reward one clear path
Google's Devpost customer story says Gemini 3 Hackathon had more than 35,000 developers, a $100,000 prize pool, and Globot won Grand Prize. The same story says Google Cloud x MLB Hackathon had more than 7,600 developers, with Jake DiBattista winning Overall Grand Prize for MLB Pitcher Mechanics Scorecard.
Those are large fields, but the winning submissions still map to clear paths. Globot handled supply-chain crisis work with multiple agents. Pull the Pitcher analyzed pitcher mechanics with Vertex AI Gemini Pro Vision. The scope is big enough to matter, then narrowed enough to demo.
Local and open models are now credible
OpenAI's winners page names RoboChef Best Overall in the OpenAI Open Model Hackathon 2025. The project is a kitchen robot that turns natural language recipes into robotic-arm steps. That is a different signal from a web chatbot because the model output has to become physical action.
For idea mining, the lesson is simple: use local or open models when they change the story. Offline operation, privacy, latency, robotics, and edge deployment are stronger reasons than 'we wanted to use the newest model'.
Mine winners without cloning them
Do not copy Nested, ChatEDU, BlinkAI, RiskWise, or RoboChef directly. Mine the constraint: personal history to recommendation, files to study artifacts, alternative input to speech, messy signals to risk triage, natural language to robot actions.
A clean Devpost winner pattern has five pieces: named user, painful workflow, visible before state, working output, and submission artifacts that match the rules. If your idea has those, it can be small and still read serious.
< read by a human · updated as things change >
browse hackathons