Skip to content
Pratik Rai
building in the open

Nagpur, India

About me

a little about pratik

I build local-first AI agent infrastructure.

Open to 2027 new-grad and internship roles in AI systems and backend engineering.

I got into computer science because of video games. That's the whole origin story. I wanted to know how the thing on screen actually worked, and once I started pulling at it the love story began. Since then I've pretty much been destined to be a nerd.

The journey wasn't straight. High school, then the actual engineering, then whatever caught my attention next: full-stack apps, cross-platform development, a long stretch in blockchain, and now scalable AI systems. I never had a five-year plan, just a habit of chasing whatever I couldn't stop thinking about.

Everyone says I'm excellent, and honestly, they have a point. I'm a genius, allegedly. A born genius, or so I'm told. But what actually separates me is the love: for computer science, for studying, for picking up knowledge just because it's interesting. And not just picking it up. I like passing it on.

Off the clock I'm tinkering, reading, or watching a movie. I play a lot. I follow cricket, football, and F1, and I'll shout at my screen for five players in a country I've never visited, hammering away at a keyboard. I like people who are as weird as I am. I chair the coding communities at my college, organize national hackathons, and I'll argue about a spicy topic until one of us gives up.

Memories

Moments that stuck.

Photos that don't fit on a resume.

Hackathon wins

What I've won.

Six wins and a national finalist finish, all shipped as working demos. Hover a panel to open it.

Principles

What I optimize for.

01

The runtime is the product.

A demo is easy. Keeping it working when the network drops or the model changes is the actual work, and that's the part I optimize for.

02

Local-first is trust-first.

If the software can't own its data, it doesn't own its future. I default to on-device, portable formats, and backends you can self-host.

03

Models recommend; rules decide.

LLMs are good at ambiguity and bad at accountability. I let the model recommend and keep the actual decision in plain, auditable code.

04

Ship working systems, not winning arguments.

I'd rather have a small, correct subsystem running tonight than a perfect architecture doc waiting for consensus.

Journey

How I got here.

Y

Yeshwantrao Chavan College of Engineering

B.Tech. in Computer Science Engineering (IoT Specialization) · Jul 2023 – Jun 2027 · GPA 7.8/10

Relevant courses: Advanced DBMS, Data Structures & Algorithms, Distributed Systems, Object-Oriented Programming, IoT Systems.

J

Junior Software Developer

DIMO — Roro Birds & Zebra Swap · Apr 2025 – Oct 2025

Web3 automotive data platform building NFT tournaments, L2 validation, and a cross-chain DEX aggregator.

T

ThinkDiff-SLM — Masked Diffusion Language Model for Edge AI Deployment

IEEE · 2025

A small language model using masked diffusion for efficient on-device inference at the edge.

W

Winner — TrackShift 2026

Recognition

Mphasis × TGR Haas F1 Team innovation challenge, 180+ students. Built an F1 telemetry replay with a hierarchical AI race engineer.

I

IEEE-published

Recognition

ThinkDiff-SLM, a masked-diffusion language model for on-device edge deployment.

H

Hackathon winner — 5×

Recognition

Enduraverse (₹1,00,000), BNB Hack Kerala, HackGenX (₹50,000), SVPCET Hack-A-Thon (₹50,000), L&T Createch 2025 (₹50,000). SIH 2025 finalist; finalist in 15+ national events.

Stack

Tools I reach for.

Languages

  • Python
  • TypeScript
  • Rust
  • Go
  • C++
  • Swift
  • Solidity

AI & agents

  • Multi-agent systems
  • Agent runtimes
  • LLM orchestration
  • RAG pipelines
  • Vector databases
  • Ollama

Frameworks

  • FastAPI
  • React
  • Next.js
  • Node.js
  • React Native
  • SwiftUI
  • Tauri

Infrastructure

  • Docker
  • PostgreSQL
  • Redis
  • Git
  • Linux
  • Modal
  • Cloudflare

Open source

Merged contributions.

mem0

Python · LLM memory layer

Merged a fix resolving an edge case in memory retrieval, mapping entity parameters to filters in the GET /memories API.

OpenClaw

TypeScript · agent orchestration

Merged fixes to Ollama model discovery and the config pipeline, enabling support for additional LLM providers.

Open Design

TypeScript · design tooling

Merged fixes for blocked PDF exports across the web and desktop applications.