Tapti Labs

Understanding AI, ML, software & systems — through practical builds, workflows, and clear explanations. No hype.

[Tapti Labs @TaptiLabs]

1/2 You don't need a vision model to ask a question, only to read the page. That split is worth stealing for any retrieval system with expensive documents and cheap queries.

[Tapti Labs @TaptiLabs]

Documents are visually complicated. Questions about them usually aren't.

NanoVDR exploits that asymmetry: a big 2B vision model indexes your PDFs once, offline — but a tiny 69M-parameter, CPU-only text encoder handles every query afterward, at 51ms, retaining 95% accuracy.

[Tapti Labs @TaptiLabs]

A useful rule in AI engineering:

Don’t start with the model. Start with the failure mode.

Ask:
• What can go wrong?
• Can rules solve part of it?
• Where does ML add value?
• How will we know it works?

Good AI starts with good problem framing.