Bring a sledgehammer to crack a nut. Take a problem that doesn’t need AI at all and solve it with far more AI engineering than anyone asked for. A multi-agent council that debates whether your toast is done. Retrieval over your houseplant’s feelings. A full eval suite for an AI that grades dad jokes. The more trivial the problem and the more rigorous the solution, the better.

Requirements:

• AI must do real work in the project. Hardcoded or faked responses are disqualified. The joke can be that a single if statement would have solved it, but your system can’t be an if statement in disguise.
• The project should be delightful.
• Your demo must show how the system works, not just the punchline.

What a strong entry looks like:

• Picks a problem so small that the effort spent on it is funny on its own.
• Uses real engineering techniques like agents, tool use, retrieval, evals, fine-tuning, or orchestration, and uses them well.
• Works live during the demo.
• Commits to the bit everywhere. The README, architecture diagram, benchmarks, and commit messages all treat the problem as though it were critically important.
• Makes the judges want to try it themselves.

What we mean by overengineered:
Overengineering is the gap between how much the problem matters and how much work went into solving it. Adding more parts isn’t enough on its own. A tangled pile of tools that barely runs is just a mess. A clean, well-built system that solves nothing important is the goal.

Prize: 1-month memberships to 757MakerSpace, awarded to each member of one winning team, maximum of three. Teams with more than three members decide among themselves who receives them.