Use the sponsor stack when it is the track
Microsoft's AI Agents Hackathon showcase named RiskWise Best Overall. Its visible stack included Python, React/Next.js, Azure AI Agent Service, Semantic Kernel, and SQL for a supply-chain risk analyst.
That is the right kind of stack choice: it proves the sponsor tech in the core workflow. If the rubric rewards Azure agents, do not hide Azure behind a generic chatbot demo.
Local AI is now a real lane
RoboChef won Best Overall at the OpenAI Open Model Hackathon 2025. Its project page says it used gpt-oss, Ollama, Isaac GR00T, an SO-100 arm, a Hugging Face dataset, LeRobot, and Python to run a kitchen robot workflow locally on one RTX 4090.
That stack works because local inference is part of the story. If your edge or robotics demo depends on low latency or offline behavior, gpt-oss plus Ollama can be more convincing than another hosted API wrapper.
Next.js is fine when the product needs pages
ChatEDU won the Microsoft AI Classroom Technical Project Prize with Azure Cosmos PostgreSQL, Azure AI Search, Azure AI Document Intelligence, Azure AD, Azure Static Web Apps, React, Next.js, and GPT-4. The stack matched an education workflow with files, study sessions, and retrieval.
Use Next.js when judges need a product surface: upload, review, ask, edit, export. Do not spend the weekend building a design system. Use the framework to make the working path obvious.
Supabase patterns fit fast data apps
Crocodile, an ETHGlobal Agentic Ethereum showcase project, used Next.js, shadcn, Hyperlane, Flow, Polygon Amoy, Sepolia, and Supabase for a cross-chain supply-chain monitor with AI chat insights.
That is a useful pattern for hackathons: managed auth, data, and UI primitives around a specific workflow. If your idea needs accounts, rows, files, and a demo dashboard, Next.js plus Supabase is usually enough.
Agent stacks need visible state
Globot won Gemini 3 Hackathon Grand Prize with Python, FastAPI, CrewAI, React, TypeScript, Deck.gl, Gemini 3 Flash, Gemini Vision, Gemini Embeddings, Clerk, SQLite, and ChromaDB. Its project framed agents around supply-chain crisis management, maps, compliance, and human-in-the-loop decisions.
For agent projects, show logs, intermediate decisions, tool calls, and fallback paths. Judges cannot score invisible autonomy. A simple state table beats a complex agent that only returns a final paragraph.
< read by a human · updated as things change >
browse hackathons