About

Tanmay.

Software engineer in Mumbai. I build backend systems and advise clients on the side.

I work full-time at GT Bharat writing Go and TypeScript — APIs, distributed systems, the kind of backend work where the interesting problems are usually about tradeoffs nobody documented. On the side I advise clients on architecture decisions and help engineering teams think through technical strategy.

The combination is unusual. Most engineers don't do client work. Most consultants can't ship code. I do both, and the overlap is where I think most clearly — I know what's technically real, what's technically feasible, and how to explain the difference to a room that doesn't.

The writing

Most technical writing falls into one of two traps — too shallow to be useful, or too academic to be actionable. I'm trying to write the middle version.

Opinionated takes on AI and systems design, grounded in what actually works in production. When I cover a model or framework, I go deep enough that you could make a real decision based on it. Not a summary of the press release. An honest read.

Every post has architecture diagrams — SVG charts built to show the system, not decorate it. The topics follow what I'm actually paying attention to right now: multi-agent systems, LLM orchestration, and the gap between research demos and production deployments.

Sakana Fugu
Multi-agent orchestration without a big model
CrewAI vs LangGraph
Mental models, not capability
Constrained decoding
Why prompting for JSON fails 1 in 10 times
Multi-agent jailbreak
Attack surfaces nobody is watching
LangGraph in production
From invoke() to WebSocket streams
EPIC framework
Communication as a system design problem

How I write

Research first. Primary sources — papers, changelogs, arXiv — not secondhand summaries. If a number appears in a post, it has a source.
Diagrams that explain. Every post with a system in it has an architecture diagram. SVG, dark terminal theme, built to show the structure not dress the page.
Honest benchmarks. If a model trails on some tasks and leads on others, I say both. "Shoulder-to-shoulder" is marketing. Selective parity is the data.
No filler. I publish when I have something worth saying. Roughly once or twice a month.

Get in touch

For consulting. I take on a small number of projects at a time. [email protected]
For the writing. Follow on GitHub or subscribe via RSS.
Portfolio. tanmaybohra.com ↗