I've been writing code since a school computer lab around 2000: BASIC, then Java, then C++, then whatever the next thing needed. These days I build products that use language models. Most of that work is keeping them honest: limits they can't override, checks on what they say, and knowing what each call costs.

More about me, and my family, at lcbasu.github.io.

• Proofsource: measures what ChatGPT, Claude, Perplexity and others say about a brand, with error bars. Also as plugins for Claude, Claude Code, ChatGPT and Codex.
• Pagerless: on-call for apps. Agents try the fix first; a person is paged only when needed. Go SDK.
• Rivane: an accounting OS on top of ERPs, which I co-founded.
• Driffle: AI meeting notes, with a Rust core.
• g4a (repo): keeps an AI coding agent's reasoning in git, so you can see later why the code changed.

• Fast Mail Tracker: Gmail tracking, as a Chrome extension and web app. Passed Google's CASA security assessment.
• Visitor Signal Engine: one script tag shows which companies visit your site. Runs on Cloudflare's free tier.
• claude-chrome: use Claude in Chrome with several Claude Desktop accounts on one Mac.
• Promptset: a macOS menu bar prompt library that builds itself from what you reuse. Fully local, with no network access. Not public yet.
• homebrew-tap: Homebrew formulae for the above.

small-models-vs-frontier: small open models fine-tuned on one Mac against Claude Sonnet 5 on three tasks. One win, two ties, 30 to 51 times cheaper, $7.83 in API calls. Every call is cached.

Most of the repos further down are from 2012 to 2018, and I've left them up on purpose: Khai-Khai, a game from a Microsoft hackathon at IIT Roorkee; graphics and VR playgrounds from the SmartVizX years (Unity, Unreal, FBX SDK, OpenGL, ReactVR); my semester notes from IIT Roorkee; and a lot of Rails, Android and iOS practice. Products from those years that are no longer live: SuperDev, Unbox, DukaanKhata and MyWasiyat.

Find me on LinkedIn.

On X I mostly retweet other people.