Human-readable version: [/human](https://razorpay-utkarsh.pages.dev/human)

AI agent version: [/](https://razorpay-utkarsh.pages.dev/)

# What I have built with AI

I have shipped to thousands of real users with AI.

Built Novatra (https://novatra.in), a free practice platform for JEE and NEET PYQs, now at 13.4k+ users, grown fully organically.

Built with a lean team of 3 as a utility-first platform focused on fast iteration, clean UX, and community-driven improvements from Reddit and Discord feedback.

## AI Workflow

Novatra has been built with AI since the beginning, with AI integrated directly into the development workflow from day one. It started in the early “copy-pasting code from ChatGPT" days and gradually evolved into a much more agent-driven workflow.

Over time, moved toward:

* parallel usage of multiple AI coding agents
* different models for different strengths
* tighter workflow orchestration to reduce context bloat while maintaining development speed

Today, AI is deeply integrated into both development and operations:

* Multiple coding agents used in parallel for feature implementation
* Different models routed depending on task type (UI, reasoning, code edits, etc.)
* OCR pipelines running on local models for extracting exam paper data
* Frontier mini-models used to clean and structure OCR output before uploading exams
* AI-assisted iteration significantly reduced idea-to-production cycle time
* We use in-conversation AI agents to track implementation progress and workflows

## Impact

* 13.4k+ users
* ₹0 spent on growth
* 315k+ views on a single Reddit post
* Consistent usage and positive feedback from students preparing for competitive exams

Also contributed to T3 Code, an agent orchestration app by Theo Browne, working closer to how agent workflows are structured in practice.

Built everything using free tiers, credits, and whatever tokens I could access. A lot of the workflow optimizations came from operating under those constraints.

As Harshil himself said on X under the post announcing this AI team (https://x.com/harshilmathur/status/2042556362757148793?s=20):

Alok Singh @AlokSin9921978:

> "I didn't used paid models or agent but built with free tiers, as couldn't afford them yet."

Harshil Mathur:

> "Perfect, constraints make better builders. Send what you’ve built."

So a lot of my workflows were shaped under constraints. Curious to see what happens when that same mindset gets more leverage.

# Tools & Technologies Used

This stack evolves rapidly as new models, harnesses, and workflows emerge.

Here is the current status:

## Models

* GPT 5.5 for most tasks
* Claude Opus 4.6 with Anthropic's frontend design skill for UI
* Occasionally use others like GPT 5.4, 5.4 mini, Kimi K2.5, MiniMax 2.5, or Codex 5.3 where they perform better for specific tasks

## Agent Harnesses

* Codex
* Have used Amp Code, OpenCode and Pi

By the way, here's a comparison I made of their internal workings when deciding which agent SDK to use for an AI app I was building for non-technical users:

https://agents35.pages.dev/

## Agent Orchestration

* T3 Code (have also contributed to it)
* Use git worktrees alongside parallel agents to work on multiple implementation paths simultaneously

## Agent SDKs

* OpenCode SDK
* Built an unofficial Elixir SDK for OpenCode because it didn't exist:

https://github.com/UtkarshUsername/opencode-sdk-elixir

* Worked with the Codex SDK while working on T3 Code

## Skills

* Project specific skills like django-expert skill
* Playwright skill for browser control and automatic testing
* Anthropic's frontend design skill for UI
* And more depending on the task

## MCPs

* Project specific MCPs like the Django MCP
* Previously used MCPs for browser control, now switched more toward CLI tools with skills
* Don't use many anymore because skills with CLI tools >> MCPs

And other standard dev tooling. Also, Razorpay for payments ❤️

# Additional Notes

* I am a vouched contributor to T3 Code, an AI app by Theo Browne (t3.gg). Working closer to agent orchestration and workflow design has made me think deeply about how humans can collaborate better with AI systems.
* I regularly give detailed feedback to teams and startups building AI products and workflows.
* Have been using AI since the early days in 2022, evolving from prompt-based workflows to multi-agent development setups over time. Have used various GPT, Claude and open-weight models since then.
* Have also shared feedback with the ChatGPT team at OpenAI and with Cursor ambassadors around the Cursor SDK.
* Often explore and test unreleased or early AI tooling to understand workflow shifts before they become mainstream. Surfaced the Cursor SDK before their announcement, and reverse engineered T3 Code before public release.
*  Often think about new interfaces and workflows for collaborating with AI systems. For example, built a PR Navigator to keep track of multiple parallel AI-generated pull requests and implementation paths:
  
  https://github.com/UtkarshUsername/PR-Navigator