About vakforge

vakforge is a pre-alpha CLI and coding-agent skill for developers building self-hosted voice assistants from their own data. It helps answer an early question: does this project need retrieval, tools, locale support, or fine-tuning—and what evidence is missing?

• Inspect: Profiles documents, tables, chats, and recordings. It identifies possible lookup tools and personal data, records files it could not read, and measures audio properties. It does not transcribe calls.

• Recommend: Gives a decision for each goal, with the evidence, data gaps, and reasons behind it. It can recommend starting with retrieval and tools instead of training.

• Report and validate: Produces a shareable HTML decision report and checks canonical datasets for schema, provenance, consent, and redaction requirements.

• Agent skill: Guides a coding agent through the later preparation, training, evaluation, and serving stages. This full path has not yet been tested end to end on a real project.

The current CLI runs locally, with no ML framework required for its core commands. Locale packs currently cover English (US, UK, India) and Hinglish. The project is Apache-2.0 licensed.

Current status: init, inspect, recommend, report, and validate work today. Training recipes, end-to-end evaluation, and serving remain planned. Feedback from developers and ML engineers is welcome, especially on the recommendations and real-world data formats.