Your first week on ML/Research. Do the company onboarding first, then this.
What we’re solving
Honcho’s differentiator is the modeling layer — predicting an entity’s state, intent, and preferences, not just storing facts. Read what we optimize for the framing, then the research reading list starting with the four areas of ML for honcho.
The ML pipeline
The flow is benchmark → trace → train → serve:
| Stage | Project | Does |
|---|---|---|
| Benchmark | Excadrill | Boots Honcho harnesses, runs eval benchmarks |
| Trace | Minccino | Turns reasoning traces into SFT/DPO/GRPO datasets |
| Train | Machamp | Fine-tunes models (Unsloth + Hydra) |
| Serve | Metagross | vLLM inference for the Tinybox |
System Map shows how this connects back to Honcho.
Hardware & access
- Porygon (Tinybox, 6×4090) and Varoom (DGX) — see Infra. Both are Tailscale-only, no public IPs.
- Get on Tailscale and ask a founder for access to the boxes.
Evals
- DRIFT is our internal benchmark for modeling quality (still in design).
- Public benchmarks (LongMemEval, LoCoMo, BEAM) are table stakes — run via Excadrill.
Where to start
- Read Deriver Training Pipeline for repository ownership, artifact handoffs, and the release process.
- Run one benchmark end-to-end with Excadrill.
- Read the Honcho ADR so you understand the data the models see.
- Log any snags in the friction log.
TODO (needs a founder/ML lead to fill in): current training datasets, active research questions this quarter, and who owns each pipeline stage.