Team 51

tuno

tuno is an autonomous training agent. Describe a behaviour; it recalls past runs (Cognee), finds and reranks licence-clean data (context.dev + Superlinked), ingests it (Airbyte), quotes and rents a GPU (JarvisLabs/modal/e2b), runs a live LoRA fine-tune, and keeps a receipt for every cent.

Check your own project

Community votes recognize the build and do not affect official judging.

Review the project

Start with the source code, then open the demo or video if available.

View GitHub repository
Project description
Most AI tasks no longer need a frontier model: a small fine-tuned model does the job better and cheaper. tuno is a personal agent that does the whole job for you. How it works (every step is a real call, or is labelled replay/fallback): 1. Cognee: cross-context memory. Before quoting, the agent recalls lessons from past runs (e.g. its first cold GPU start cost 1.84× the estimate), so it reuses warm GPUs and corrects its quotes. It's a self-improving loop. 2. context.dev searches for datasets. Superlinked SIE (Qwen3 reranker) scores each candidate against the goal. 3. A licence gate keeps CC BY-NC human-transcribed Expresso and rejects unlicensed, pseudo-labelled ASMR. 4. Airbyte (PyAirbyte OSS, source-file → DuckDB) ingests 1,139 whispered clips. You can play them in the UI. 5. A live gpu provider price feeds a quote corrected by memory. You approve. 6. Dispatch: a Boat sandbox (any provider of choice) control plane warm-resumes the GPU and runs whisper-small LoRA SFT. It streams loss and WER every 15 s to a live Jobs view and to Amplitude, Airbyte syncs the GPU log into the audit store, and the GPU auto-pauses with hard caps. Result today: word error rate on whispered speech went from 7.83% to 6.95% for $0.05 of GPU time, with zero GPUs left running. The audit store is SingleStore-ready (SQLite fallback). Next: any task, SingleStore live analytics, a Cognee knowledge graph, and computer-use models.
Tools used