Work I designed, built and still run

The problem

A five-minute roll holds dozens of transitions, scrambles and positions that look alike. Coaches can't rewatch every round for every student, so detailed feedback rarely happens. Can a multimodal model act as a reliable second set of eyes for an instructor, and turn that feedback into the next class?

It's also a hard benchmark for video AI: long footage, fast action, and a half guard that looks almost the same as a quarter guard. That's why TKNQ keeps CodeJitsu running as its R&D testbed.

Design decisions

• Analysis runs as a background job (Hangfire on Postgres), with progress pushed to the browser over SignalR.
• A person stays in the loop. Instructors can edit every AI output before a student sees it.
• Model choice is a setting, not code. An admin AI Console sets the model, region, reasoning level and token limit per feature, and every run snapshots its settings, so a new Gemini generation is a configuration change instead of a redeploy.
• Knowing when not to use AI. Technique search first used Grok live search (dead links), then Gemini with Google Search grounding (too slow for the server's time limit). A plain YouTube Data API search replaced both: instant, cheap, and only real videos.

Running a paid AI feature as a free demo

Every analysis costs real money, so the controls had to be fair to users and safe for the budget:

• 3 analyses per account per rolling 24 hours, adjustable live from the AI Console.
• Failed runs are refunded automatically, and automatic job retries are off. Before that change, one failure could quietly re-run the whole paid analysis up to ten times.
• The limit is checked before upload, two requests can't both take the last slot, and a double-click never costs two runs.

Built with a team of AI agents

Since March 2026, CodeJitsu has been developed with Claude Code working as five specialised agents, each defined by a role file in the repository. I set direction, answer the product questions and review what ships.