Now

This is a now page. Updated 17 September 2026, from San Francisco.

An autonomous org for Kalamandir Jewellers

I'm building Aera, an operations brain for multi-unit operators, and the project I'm on now is putting it to work inside Kalamandir Jewellers: a jewellery house founded in 1986 that did around ₹5,850 crore (roughly $700 million) in FY25, with flagship showrooms across Gujarat and Mumbai, airport outlets around India, and its own lab-grown and omnichannel brands.

A retailer that size runs its buying on legacy systems and people's memory. Purchase orders, vendor and manufacturer coordination, and stock moving between showrooms, with an enormous amount of working capital tied up in gold inventory. So the wedge is the vendor side. Aera sits on top of the systems they already run, no rip-and-replace, and starts by taking over the procurement and vendor-facing work: reading vendor messages, drafting and reconciling purchase orders against real stock movement across stores, and surfacing the exceptions a person should see.

It works the way everything else I build works. The model proposes, code enforces the policy, a human approves anything consequential, and every action lands on an audit trail. The owner gets one daily brief and approves from chat. Prove that one wedge, then grow it into the broader autonomous org, where more of the operation runs itself as it earns the trust to. An early version is already running on the group's real data.

Talking to a few early-stage teams

About founding-engineer and forward-deployed roles where the job is agents inside a real system of record. Two technical case studies from those conversations are public: Meridian OS and Carrier Desk.

Drilling fundamentals

Live coding without an agent, from scratch, on a timer. I build with AI agents daily, and I want the hand to be as strong as the judgment. The specific thing I am drilling is state selection: what state and what data structure a problem actually needs, decided before the first line.

Reading

Whatever is actually moving agent reliability forward, which right now is reinforcement learning in post-training and honest evaluation, not longer prompts. I read Nathan Lambert's Interconnects and Lilian Weng for the research, Anthropic's and OpenAI's engineering write-ups the day they land, and the steady stream of agentic-RL, on-policy-distillation, and long-horizon-memory papers coming out of the frontier labs. Under all of it is the question I keep running into in my own work: where does trustworthy behavior come from, the model or the system around it, and how much of it can you actually measure.

Not doing

Podcasts, conferences, or content for its own sake. If I write, it is because a repo needs its argument written down.