The Flywheel
The flywheel is Ody's core thesis: every product feeds the next, creating a self-improving cycle of knowledge quality and AI accuracy.
How It Works
User runs Refine → finds contradictions → resolves them → training data created
↓
Forge takes training data → trains custom model → evals → deploys
↓
Colleague uses custom model → answers questions → user corrects mistakes
↓
Corrections become new training data → back to Forge → model improves
↓
Repeat forever. Every correction is an investment.
Step by Step
| Step | Command | What happens |
|---|---|---|
| 1 | ody-refine ./docs/ | Health report + findings |
| 2 | ody-refine resolve | User triages findings, creating preference pairs |
| 3 | ody-refine export --format trl | JSONL training data |
| 4 | forge train from-refine --method auto | SFT then DPO on a base model |
| 5 | forge eval | Before/after comparison, prevents regression |
| 6 | forge deploy | Push model to registry |
| 7 | Colleague serves the model | AI assistant answers with domain knowledge |
| 8 | User corrects wrong answers | New preference pairs generated |
| 9 | Automatic retrain trigger | Model improves from corrections |
Why It Matters
Each correction makes the model more aligned to how your team actually thinks. The longer you use Ody, the better it gets.
The 30-Second Pitch
"When you run Refine, it finds every place your company's knowledge is broken. When you fix those -- marking which version is correct -- that decision becomes training data. Forge trains a model on those decisions. You own that model. Every correction makes it more aligned to how your team actually thinks."