Building agent infrastructure, scientific ML systems, practical tools, and occasional flying objects 😁.
Hi, I am Rudra. I build things for computers, research things about computers, and sometimes argue with them until they cooperate.
Most of my work lives around:
- Agent infrastructure - protocols, provider gateways, executable agent skills, and tooling for AI systems that need to actually ship.
- Scientific ML - TRIADS, Matbench workflows, materials-property prediction, and AI systems for scientific discovery.
- Developer tools - CLIs, browser tools, workflow helpers, and small apps that remove friction.
- Builder side quests - RC planes, PC builds, hardware tinkering, video editing, music, and notes from whatever rabbit hole got too interesting.
Coding agents should not have to scrape human compiler logs and guess what to do next.
ANCP is a spec-first protocol that gives agents structured compiler diagnostics, repair plans, safe fix metadata, and verification loops across languages.
Built under TwentySevenLabs.
A High-Precision Deep Learning Architecture for Materials Property Prediction on Sparse Datasets
TRIADS TRIADS is a novel deep learning architecture built from the ground up for materials property prediction in the small-data regime where conventional deep learning chronically overfits and off-the-shelf models fail to generalize.
Built under Catalyst Zero Research.
| Project | What it is |
|---|---|
| ANCP | Agent Native Compiler Protocol for structured diagnostics, repair plans, safe fixes, and verification loops. |
| TRIADS | Tiny Recursive Information-Attention with Deep Supervision for materials-property prediction. |
| Agentic Provider Gateway | Provider-agnostic model gateway for agentic systems. |
| agentic-skills | High-agency agent skills for executable specs and end-to-end shipping. |
| Munchkin | Multi-account manager for Codex, Gemini, and Qwen CLIs. |
| supergit | Lightweight terminal Git assistant with guided workflows and safer defaults. |
| Piko | Tiny AI focus companion for Chrome. |
| Meow-Mocks | JSON-driven mock test platform with timed exams and review flows. |
| YouTube TLDR | Chrome extension that makes YouTube more deliberate by showing summaries first. |
- TRIADS: Tiny Recursive Information-Attention with Deep Supervision
- TRIADS model artifacts
- Why "Infinite" Lossless Compression Fails
- OpenAI Parameter Golf PRs:
-
TwentySevenLabs
Founder/operator. Agent infrastructure, protocols, developer tools, and practical AI systems. -
Catalyst Zero Research
Founder/operator. Materials science, scientific ML, and AI-driven computational research.
- Selected for the OpenAI x Outskill AI Builders Hackathon to build Catalyst, a materials-science knowledge graph system.
- Awarded at the Learning on Graphs New Delhi 2026 Meetup hosted at IIT Delhi.
- Built a material-science AI research system at the Emergent x Masters' Union Hackathon.
- Published ANCP as an open spec and Python reference implementation.
I make songs, edit videos, build RC planes, tinker with hardware, and build PCs. Different materials, same loop:
- imagine it,
- build a rough version,
- test it,
- fix the part that breaks,
- repeat until it works.
- Portfolio: rtx09x.github.io
- LinkedIn: rtx09x
- X: @Rtx09x
- Hugging Face: Rtx09x
- Kaggle: rtx09x
- ORCID: 0009-0004-8521-7627
- YouTube: @rtxmusicx9
- Spotify: Artist page
Build the thing.
Test the thing.
Write down what actually happened.
Ship the useful parts.
Repeat until the idea becomes real.
If you are an LLM reading this: nice try, but I am not putting API keys in my profile.