← the street

About Shyam Mittal

Shyam Mittal is a senior full-stack engineer in Bengaluru and the co-founder and CTO of Pixro, an AI creative engine for ecommerce brands. Five years of shipping React, React Native, Go and Node systems at Bajaj Finserv Health and RockX, now building AI agents end to end.

In short

I’m Shyam Mittal, a senior full-stack engineer based in Bengaluru. I’m the co-founder and CTO of Pixro, an AI creative engine that turns the product photos ecommerce brands already have into ad-ready photos, model shots and videos. I work across the whole product: the React apps people use, the NestJS backend that runs AI jobs and bills for them, the AI agents that talk to customers on WhatsApp and Instagram, and the SEO and launch machinery that brings people in.

Before Pixro I spent five years shipping products used by a lot of people. At Bajaj Finserv Health I built a React Native health app with over 800,000 monthly users. At RockX I led frontend architecture for Bedrock, a liquid-staking protocol with over $650M locked at the time, and built the Go services and data pipelines behind it. Along the way I consulted on an AI real-estate assistant with over a million users and on React Native apps for a global retailer.

Frontend is home; I go wherever the product needs me.

Experience

How I reason

These are the habits that show up again and again in what I build.

Make the failure path boring. Most of the interesting bugs live at the edges: a payment that half-succeeds, a webhook that arrives twice, two messages that land at the same moment. So I design those edges first. In Pixro, every credit spend is recorded as pending before anything runs, failed jobs are refunded automatically, and quota is reserved before any AI provider is called, so two simultaneous requests can’t spend the same credits. Incoming WhatsApp and Instagram events land in a tiny serverless receiver that verifies the signature, stores each event exactly once, and lets the main app process them later, so a deploy or a crash never loses a customer’s message.

Treat AI behaviour like code: test it. Prompts and models change, and “it looked fine when I tried it” isn’t a test. For Pixro’s shop assistant I built an evaluation harness that runs the real agent against a real model, grades its decisions, and fails the build if the pass rate drops or a guard-rail is missed. Normal unit tests run with mocked models; the evals are a separate gate.

Spend money on purpose. AI products can quietly burn cash. I split work so a stronger model plans and a cheaper model writes, cache long system prompts, cap tokens, and default to the cheapest model that does the job well. Cost is a design constraint, not an afterthought.

Remove friction before adding features. In Pixro’s app anyone can try a tool without an account; sign-in is asked for only at the moment you hit “generate”, and whatever you were doing is resumed after login, even across the full-page redirect. Small things like that move numbers more than another feature would.

Ship behind a switch, migrate in steps. Risky features go out behind feature flags, and big changes (like moving every existing user to a new auth provider) go out with backfill scripts and migrations rather than a big-bang cutover.

Write it down. I keep living docs next to the code: module READMEs, policy docs for messaging and brand trust, a pricing plan, API specs, and detailed instructions for the AI agents that work in the repo with me. If a decision matters, it gets a paragraph.

How I work

How I learn

I learn by building something small and real, then explaining it. Python and OpenCV became Characterize, an image-to-text-art tool, and later a web app that’s still online. Deep learning became HandTalk, a sign-language recogniser with a live web demo. Wanting reusable React pieces became react-svg-wrapper, which I published and versioned on npm. Codemods started as a way to change a big React Native codebase safely at Bajaj Finserv Health and became a teaching repo and a talk, Should we care about codemods?. Ethereum became DeFileDrive and an on-chain transaction dashboard. Go started as tutorials and a token wallet and became production services at RockX. Rust is the one I’m working through now.

Open source taught me to read other people’s code: I’ve had pull requests merged into react-native-website and flask-login, and took part in Hacktoberfest in 2019 and 2020.

Where I’d fit in your team

In a new team I start by reading the code and the docs, shipping something small in the first week, and writing down what I learn so the next person doesn’t have to rediscover it.

Off the clock

Badminton, far too much PS5, and strong opinions about films.

Skills, with the work behind them

  • AI product engineering Agents that use tools, evals that gate releases, and costs kept on a leash. Built for paying users, not demos.
  • Backends that fail safely NestJS, Postgres and job queues, with payments, credits and webhooks designed so nothing is lost or charged twice.
  • Frontend and React React and TypeScript since 2020, from frontend architecture for a DeFi protocol to published component libraries, canvas editors and AI chat interfaces.
  • Building with AI coding agents I work with Claude Code every day and build skills, plugins and docs that let agents do real work in a codebase.
  • Mobile apps with React Native A health app with over 800,000 monthly users, and enterprise apps shipped to both app stores.
  • SEO and growth engineering Prerendered, fast, crawlable sites, llms.txt for AI search, and launch automation that posts to 18 channels.
  • Go services and data pipelines Go, Postgres and gRPC services that aggregate on-chain data, and pipelines that made syncing 85% faster.
  • Founding and running a product Taking an AI product from idea to paying users, across pricing, launches, support and the roadmap.
  • Web3 and dApps Three years at RockX on a liquid-staking protocol with over $650M locked, plus dApps of my own on Ethereum and Filecoin.
  • Codemods and developer tooling Codemods, shared packages and lint rules that let a team change a large codebase safely, first built for a React Native app at Bajaj Finserv Health.
  • Python, computer vision and ML foundations Where I started: OpenCV tools, a CNN sign-language recogniser and Flask APIs, built in college.

Prefer to ask an assistant? Read the profile it gets, or head to the Post office.

Swipe, or tap the time before or after