Start from a workflow, not a model
AgentHacks 2025 tracks included agentic systems and workflow automation, human-AI collaboration, personalization and memory, and AI safety and control frameworks. Those tracks are useful because they point to jobs, not just model features.
Good weekend ideas start with a repeated task: triage, search, summarize, route, verify, plan, or hand off. If the model call is not attached to a workflow pain, the project will read like a demo of someone else's API.
Agents win when the job is specific
Microsoft's AI Agents Hackathon named RiskWise Best Overall. RiskWise is a supply-chain risk analyst that flags disruptions from shipping, news, and data, then lets analysts query risk. WorkWizee, another Microsoft category winner, updates Jira, Bitbucket, Confluence, Outlook, and Teams from incident-call language.
Derived weekend ideas: supply-chain watcher for one SKU, incident-call to ticket updater, grant eligibility checker, or lab inventory reorder agent. Keep the agent boxed inside one workflow with human approval at the end.
Use local or edge AI when it matters
RoboChef won Best Overall at the OpenAI Open Model Hackathon 2025 with a local kitchen robot that turns natural language recipes into executable robotic-arm steps. The project page emphasizes local execution on one RTX 4090, with gpt-oss, Isaac GR00T, SO-100 arm, Hugging Face dataset, LeRobot, Ollama, and Python.
TreeHacks 2025 also shows the edge pattern. SurviveX used on-device LLM inference for a hands-free survival assistant. A weekend version could be an offline first-aid voice guide, a local document Q&A kiosk, or a no-internet field checklist demo.
Multimodal access is a clean beginner angle
TreeHacks winners BlinkAI, Iris, Edith, and BrailleBot all turn new input or output modes into access. BlinkAI converts blinks to speech. Iris is a real-time vision assistant for blind users. Edith is framed as a second pair of eyes. BrailleBot is a low-cost braille printer under $15.
Derived ideas: blink-to-speech desk app, screen-description copilot, cheap braille label preview, or camera-based object explainer for one room type. These ideas demo well because judges can see the before and after without reading a long architecture note.
Trust and verification are product features
Microsoft's ModelProof: Sentinel AI Chat won the JS/TS category with a safety wrapper that runs dual-model checks and reports hallucination, bias, and toxicity risk. That is a useful signal for 2026 ideas: verifying AI output can be the product, not an afterthought.
Derived ideas: agent answer auditor, medical note red-flag checker using synthetic data, source-backed policy explainer, or code assistant that separates evidence from guesses. The demo should show what got flagged and why.
Use sponsor constraints as idea fuel
Google Cloud Rapid Agent Hackathon required a functional AI agent powered by Gemini and Google Cloud Agent Builder that integrates one partner MCP server. The announced prize amount was $60,000 across partner pools, so sponsor constraints were not side details. They were the design surface.
When a hackathon names a tool, build an idea that makes that tool necessary. For MCP, that could be a calendar-to-task agent, codebase helper, or CRM updater with explicit approvals. For vision, make the camera input essential. For local models, make offline mode part of the story.
< read by a human · updated as things change >
browse hackathons