Armaan Irani.
Intelligence, engineered into the product.
I'm Armaan, an AI engineer in Pune. I came to this through mathematics: a B.Sc. here in Pune, then a master's in Data Science and AI at the University of Liverpool, where I still work as a machine learning engineer on a model that finds the direction of a radiation source from a single sensor.
Today I build AI systems for real people. At Script Lanes, a medical e-learning platform that turns handwritten notes into study material, with a vision model in front and a retrieval chatbot behind it. At Nextoar, agents that research and write outreach, and computer vision for an industrial floor assistant. On my own, whatever needs an agent, a retrieval system or an automation.
I care about the seams, the places where a confident system meets a person, and I care about type nearly as much as tokens. Everything I make ships with evals and tracing, so we know it works before anyone has to trust it.
- Based
- Pune, India
- Local time
- Right now
- Learning Fusion 360 and 3D printing.
- Small parts I design in Fusion 360 and 3D print.
- Fusion 360 and 3D printing, and German on the side.
- Thus Spoke Zarathustra, Friedrich Nietzsche.
- Multi-agent systems
- RAG workflows
- LLM evaluation
- Observability
- Computer vision
- Python
- TypeScript
- PyTorch
- LangGraph
- FastAPI
- Postgres
- AWS
- Docker
- Next.js
Slowly at first, then all at once.
Listen
I start by understanding the workflow and the people in it. What happens today, where it hurts, what 'working' would actually look like. Usually one honest conversation is enough to know whether it is worth building.
Sketch
Architecture on one page, interface sketches on another, and an evaluation plan before any prompt is written. Every decision gets a line about its tradeoff so we can revisit it later without arguing from memory.
Build
Short iterations, working software from the first week, evals running on every change. Progress shows up in a staging environment, not a slide deck.
Keep tuning
Launch with monitoring, cost tracking and a feedback loop, then improve on real usage. The first version is never the last, and I would rather be around for the second.
- 01 · 18 Oct 2026Moving a business site from Vercel to Cloudflare Pages, with the domain at GoDaddyA step-by-step guide to hosting a small business's static site for free on Cloudflare Pages: why Vercel's free plan didn't fit, moving a GoDaddy domain's DNS to Cloudflare without moving the domain, the www redirect, a staging site on its own subdomain kept private with Cloudflare Access, and the problems I hit on the way. Written after doing it for my family's catalogue site in October.
- 02 · 01 Oct 2026Putting your bot on a real WhatsApp number: a Cloud API guideA practical guide to putting your own bot on a real WhatsApp number through Meta's Cloud API: what to set up in Meta and in what order, how to connect your code, and the mistakes worth avoiding. Written after taking a client's bot live in September.
- 03 · 20 Sept 2026What writing attention by hand taught meIn the summer of 2025 I wrote GPT-2 small in PyTorch, and I wrote attention four times to get there. These are the things I only understood once I had typed them: what a head is, why the mask is a buffer, and why a parameter count came out 39 million too high.
I came to AI from mathematics and ML research, which is why the unglamorous parts interest me most: evaluation, latency, cost, the empty state, the error message. A model is a component. The product is the thing.
Outside of work I'm learning German, reading Nietzsche's Thus Spoke Zarathustra and learning 3D printing. I care about type nearly as much as tokens, and it shows in what I make. It should.
Download resumeWant to talk? Say hello.
A project, a question, or just to compare notes. I read everything and write back within a couple of days.