Notes from
the workbench.
The Aditya Lab blog: plain-language notes from building Aditya AI, an assistant for Hindi, Gujarati and other Indian languages. Architecture, tool use, reading Indian scripts and checking our own work, with the measurements behind each decision, including the ones that did not work.
The two words our OCR model got wrong were right
Our reading model disagreed with the volunteers' proofread text twice on a held-out page. Both times, the printed page agreed with the model.
What is Neural Intelligence Fabric?
A direct explanation of NIF, its lifecycle, safety model and current implementation status.
Why tool calling is not enough
The difference between emitting a function call and completing real engineering work.
Building coding agents that can prove the work
A practical evidence model for trustworthy coding agents.