WRITING ARCHIVEPROJECT ARCHIVEPUBLICATIONS
Pratham Patelprathambiren2618@gmail.com@CodeWithInfernolinkedin

I’m 19. I came to the US alone at 17, after missing my dream Indian university by about one percentile on the JEE.

Before that it was Minecraft. I wanted to know how a server let me play with someone on the other side of the world. That question turned into everything else.

At 13 I built Panchayat, a WhatsApp clone. 20 users, mostly friends and family. Tiny outcome, enormous feeling. I’ve been chasing that feeling since.

At 16 I built an AI email client that shipped features before Superhuman did.

Not because I was a prodigy. Because I was ordinary, in chaos, trying anyway.

Two mirrored monoliths on a salt flat, the nearer reflecting the farther
A Lichtenberg figure branching through a block of acrylicMacro of a moth wing eyespot mimicking a predator's eye
CLEAN

Relationship-led GTM for B2B SaaS. It indexes what a company already knows, maps the warm paths into an account, profiles buyers against real ICP signals, and runs low-volume outreach that starts actual revenue conversations. It started at a hackathon. It has backers now.

Most people say “that’s a cool idea,” then move on. A few sit at 3:42 AM and try to build it. That’s the whole gap. Not intelligence, not resources, not experience.

LOKUS

Open-source markdown editor. Rust and Tauri, local-first, 10K+ downloads. Wiki links with content-aware autocomplete, LaTeX that renders as you type, themes you can rewrite at runtime. I live in it, which is the only reason it’s any good.

SELECTED WORK

Clean AgentClean Agent

An IDE where you don’t write the code — you direct the agents that do, and they share what they learn.

Clean MCPClean MCP

Your agent burns 13,000 tokens hunting one function. This finds it in one query, offline. 47 stars.

I.R.I.SI.R.I.S

Turns an iPhone’s LiDAR into a sense of touch. Blind navigation, under 5% battery an hour.

tesstess

Ten coding agents at once, each sealed in its own git worktree, and one command that finds the stuck one.

AND I TEACH MODELS TO DISAGREE WITH THEMSELVES

A model’s first answer is usually its most confident one, which is not the same as its best one. S2C splits the work in three — a Generator that answers, a Critic that attacks the answer, and a Synthesizer that rewrites it using the attack. With PPO and RAG grounding it lands a 60% relative improvement on GSM8K.

S2C framework: Generator to Critic to Synthesizer, with a critique feedback loop back into the Generator
The critique goes back into the Generator. That loop is the whole idea.

PAPERS

WRITING

ELSEWHERE

STACK

LanguagesPython, C++, Java, SQL, JavaScript/TypeScript
AI/MLPyTorch, TensorFlow, JAX, Transformers, RL (PPO, DQN), RAG, PEFT, scikit-learn
Tools & SystemsDocker, Git, Linux, Slurm, FastAPI, Flask, Next.js, Weights & Biases
Data & CloudPostgreSQL, Redis, MongoDB, AWS, GCP (Foundational)

INDEX

B.S. in Computer Science, Gannon University, Erie, PA. Expected May 2028. Still becoming.

(plates generated, not photographed; figures from the papers)