Building Agent Infrastructure & Scientific ML Systems
I build tools and research systems for coding agents, developers, and scientific discovery. Current work includes ANCP for agent-native compiler interaction, TRIADS for materials ML, and practical infrastructure for real developer workflows.
I build public systems across agent infrastructure, developer tooling, scientific machine learning, and focused product experiments. The common thread is turning messy workflows into something concrete: protocols agents can use, models researchers can test, and tools developers can actually run.
I operate this work through both my personal account and independent orgs: TwentySevenLabs for agent/dev infrastructure, Catalyst Zero Research for materials and scientific ML, and Rtx09x for experiments, apps, research notes, and fast shipping.
Away from the terminal, I am still usually building something: RC planes, PC builds, hardware setups, videos, songs, and whatever new rabbit hole is too interesting to leave alone.
Agent infrastructure: ANCP, provider routing, and executable agent skills.
Scientific ML: TRIADS, Matbench workflows, and materials-discovery research.
Builder side quests: RC aircraft, PC hardware, video edits, music, and tiny tools.
Agent infrastructure, scientific ML, and tools that are useful outside the demo.
Agent Native Compiler Protocol: an open, spec-first protocol that gives coding agents a stable machine-readable interface for compiler diagnostics, repair plans, safe fixes, and verification loops across languages.
A parameter-efficient recursive attention architecture for materials-property prediction in small-data regimes. The repo pairs code, pretrained weights, and benchmark results with a linked Zenodo preprint.
Provider-agnostic model infrastructure for agentic systems: routing, adapters, capability metadata, structured outputs, normalized tool calls, streaming events, usage accounting, and trace storage.
High-agency agent skills for executable specs and end-to-end shipping. This is the public skill layer behind clearer planning, implementation handoffs, and ship-ready agent workflows.
Cross-platform multi-account manager for Codex, Gemini, and Qwen CLIs with encrypted storage, fast switching, quota tracking, session health checks, and Codex load balancing.
Lightweight terminal Git assistant that simplifies init, commits, sync, branch switching, and undo flows with guided prompts, safer defaults, and plain-English recovery.
A public research notebook and knowledge garden for my notes on AI, compression, neuroscience, transformers, and adjacent topics that feed into longer projects and papers.
Tiny AI focus companion for Chrome. A small, practical browser tool in the same lane as the productivity experiments: lightweight, focused, and easy to try.
Work around the OpenAI Parameter Golf challenge: small language-model training, compression constraints, score-oriented experiments, and PR-frontier analysis under the 16MB limit.
Web-based mock test platform driven by JSON exam definitions, with timed exams, review flow, results breakdowns, analysis views, and configurable test content.
Distraction-blocking Chrome extension that redirects YouTube viewing into an AI-generated summary first, so you decide whether the video is worth your attention.
I use these organizations to separate public infrastructure work from scientific research work, while still keeping both connected to the same long-term direction.
Founder and operator. TwentySevenLabs is my independent lab for agent infrastructure, developer tools, protocols, and practical AI systems. ANCP and Agentic Provider Gateway live here because they are infrastructure projects meant for other builders.
Founder and operator. Catalyst Zero Research is my research umbrella for materials science, scientific ML, and AI-native research systems. The goal is to make this the clean home for future Catalyst-branded research repos and material-discovery work.
Preprints, artifacts, and research directions behind the public work.
Zenodo preprint and public artifacts for a recursive, parameter-efficient architecture for materials-property prediction in small-data settings, combining weight-tied reasoning, deep supervision, and physics-informed features.
Rudra Tiwari · DOI: 10.5281/zenodo.19200579
A 2026 Zenodo paper exploring the theoretical impossibility of infinite lossless compression and offering a clearer framework around the idea.
DOI: 10.5281/zenodo.18404002
Recent public proof: awards, hackathons, shipped projects, and research directions that show the work is active, not just listed.
Selected from thousands of applicants for a 7-day shipping-focused AI Builders Hackathon. Building Catalyst: a materials-science knowledge graph system connecting crystal structures, properties, synthesis pathways, and research literature so AI can reason across them together.
Received an award at the Learning on Graphs New Delhi 2026 Meetup hosted at IIT Delhi, in a room of researchers, professors, and industry leaders working around AI, graph learning, and materials science. The post was later reposted by Mastercard.
Built an AI research system for materials science that searches across papers, journals, repositories, patents, datasets, and web sources to surface candidate materials and structured scientific intelligence.
Published ANCP as an open protocol/spec and Python reference implementation for agent-native compiler interaction, repair plans, safe fixes, and verification loops.
I make songs, edit videos, build RC planes, tinker with hardware, and build PCs. Different materials, same habit: make the idea real enough to test.
Pop / Electronic · 2024
I build and tinker with RC aircraft. It is the same builder loop as software: design, assemble, test, crash into reality, fix, repeat.
I am into hardware, PC building, setups, and the physical side of computing. Software is better when you respect the machine it runs on.
The main places to find my code, research artifacts, updates, and creative work.
For projects, research conversations, or collaborations.
contactme.rtx@gmail.comSource code, experiments, and public project history.
@Rtx09xResearch identity, authorship, and publication records.
0009-0004-8521-7627Professional profile, current work, and public updates.
View ProfileModel artifacts, datasets, and public ML resources.
Rtx09xNotebook and benchmark-facing work.
rtx09xBuild-in-public notes and launch updates.
@Rtx09xThe central map of my public work and links.
rtx09x.github.io