Platform Overview

Ody is the platform where companies train, own, and continuously improve their own AI from their own knowledge. It is built around three products that form a flywheel.

The Three Products

Ody Refine (open source)

Job: Take messy company knowledge, find what's broken, fix it with the user, produce clean training data.

• Contradiction detection across documents
• Staleness and time bomb detection
• Interactive resolve TUI
• HTML consulting report
• JSONL export for model fine-tuning
• CI mode and GitHub Action
• 7 source connectors (Notion, Slack, Confluence, Linear, Jira, Teams, Gmail)

License: Apache 2.0. Free forever.

Ody Forge (proprietary)

Job: Take preference pairs from Refine, train a domain-specific model, evaluate, deploy.

• 2-stage SFT then DPO training pipeline
• Remote training on Modal cloud GPUs
• Eval gate to prevent model regression
• Model registry and versioning

Ody Colleague (proprietary)

Job: Be the AI assistant that knows YOUR company. Powered by YOUR fine-tuned model. Designed to improve from corrections.

• Web app with chat, voice, and mind view
• Slack bot for team-wide access
• macOS native app
• Background job processing (ingestion, distillation, promotion)
• Every correction feeds back into the training loop

Revenue: Seats + consumption.

How They Connect

Refine CLI              Forge CLI              Colleague
─── scan docs ──→       ─── import JSONL       ─── uses trained model
─── resolve ──→         ─── train model ──→    ─── user corrects ──→
─── export JSONL ──→    ─── eval + deploy      ─── new pairs ──→ Forge

The JSONL file is the interface between Refine and Forge. The @useody/* npm packages are the shared contract.

Deployment Models

Mode	Trust level	Description
Local	Zero trust	CLI runs on your machine. No data leaves.
Team	Partial trust	Colleague runs as service. Feedback flows to Forge.
Full ecosystem	Managed	Refine + Forge + Colleague. Sandbox isolation per customer.