AI kits for voice, chat, and agentic workflows: small, focused tools you can read in an afternoon and put in a product.
Most of our work explores one question: what does it take to put real-time AI (speech, agents, generated code) inside an actual product, not a demo notebook? The repos here are the answers so far, at varying stages of maturity.
| Project | What it is |
|---|---|
openbeam |
Real-time Bible verse detection for live sermons, in the browser. Streams sermon audio through Deepgram and runs four merged detection strategies: reference matching, semantic search over a 31k-verse index, verbatim quotation matching, and a confidence-weighted ensemble. Live at openbeam.tensorkit.ai. |
candy |
A specification language for stateful backends. Write actors, flows, and policies once in .candy; an LLM generates idiomatic Go, Rust, TypeScript, or Python from the spec. Experimental. |
yugo |
A GPU-accelerated interface for the Unitree Go2 robot dog: phone as controller and voice wand, cloud LLM and vision as the mind. Hackathon build. |
hezra |
A voice-powered financial assistant prototype: OpenAI Realtime voice client over a Go server handling transfers, invoicing, and virtual cards. |
Everything here is an experiment. We build to find out what holds up in a real product, and we keep the work public because the code and design notes may be useful to someone else running the same experiment, not because these are supported products. Some experiments graduate into things people use every week; most earn their keep by answering a question.
If something here is useful to you, issues and PRs are welcome on the individual repos.
