Adhiraj Singh

Kolkata, India

Adhiraj Singh

I build agent tooling, a game engine and edge perception for small hardware, and I write down what broke.

GitHub Writing

chips collected 00

Lead it with your mouse

This toy needs WebGL2, which this browser does not have.

About

I'm Adhiraj Singh. I work from Kolkata, West Bengal, India. Four rules run through most of what I build, and each one came out of something that broke.

  1. Prefer the simple component that can be checked over the "smart" one that cannot.

    TerraSight is designed so a false "safe" verdict cannot happen. It shipped without a neural network.

  2. Deterministic code does the maths; models only reason and route.

    In my Personal CFO, Gemma reasons and routes, but every number comes from deterministic Python.

  3. Freeze the contract first, then build in parallel.

    We froze the API contract on day one and built the whole thing in parallel.

  4. Measure, then delete what does not earn its place.

    My object tracker went from 4.7 FPS to 64 FPS, CPU only, by deleting the fancy parts.

Stack

  • TypeScript
  • Python
  • C++
  • Rust
  • JavaScript
  • Node
  • Next.js
  • React
  • Three.js / WebGL2
  • ONNX Runtime
  • Supabase
  • MongoDB
  • Claude Code
  • MCP

Projects

What I've built

  1. 01 TypeScript

    Weave

    Governs coding-agent runs: what agents may touch, how their work is checked, what a person must approve, and a record of every step. I built this site with it.

  2. 02 TypeScript + Rust

    engine-ecosystem

    An install-what-you-want game-engine ecosystem: math, physics, ECS, rendering, world, gameplay, networking and orbital mechanics, with a Rust numerics crate. No build step; one math core under three renderers — a WebGL2 rasterizer, an offline path tracer and a GPU raytracer.

  3. 03 TypeScript + Python

    TerraSight

    Onboard edge-AI terrain perception for autonomous planetary rovers, built for Smart India Hackathon. I designed it so a false "safe" verdict cannot happen, shipped it without a neural network, and quantized it to INT8 with zero class flips.

  4. 04 Python

    Object-Tracking

    A real-time object tracker I took from YOLOv8 + DeepSORT at 4.7 FPS to ONNX + an IoU tracker at 64 FPS. That is 13.6× faster, on CPU only.

  5. 05 Python

    Multi-Agent Personal CFO

    A local-first financial advisor where Gemma reasons and routes, but every number comes from deterministic Python.

  6. 06 TypeScript

    can-i-actually-apply (MeritMind)

    An eligibility checker for students — scholarships, internships, competitions — with a runtime trust layer over the model's JSON.

  7. 07 JavaScript

    Auto-blog

    A tiny zero-dependency static blog generator. Posts are Markdown or HTML, and it deploys to Vercel.

  9. 09 JavaScript

    devx-engine

    The one folder I clone at the start of every hackathon: skills, MCP servers and reference docs.

Writing

What broke

I write down what broke. Every post is here.

  1. I Built a Game Engine With No Build Step
  2. I Shipped "Edge AI" for a Rover With No Neural Network
  3. My Hackathon Starts With One Folder
  4. My Rover's Terrain Classifier Is Forbidden From Ever Saying "Safe"
  5. We Froze the API Contract on Day One and Built the Whole Thing in Parallel
  6. One Math Core, Three Renderers
  7. Every New Game Should Need Less Engine Code Than the Last
  8. Building a Multi-Agent Personal CFO with Gemma
  9. How I Made My Object Tracker 13× Faster by Deleting the Fancy Parts
  10. Never Trust an LLM's JSON — Building a Runtime Trust Layer

Contact

Everything I build lives on GitHub. For anything else, mail me or find me on LinkedIn or Instagram.

Reach me