01
Small models, shipped
I train from scratch, fine-tune (LoRA or full) and run RL with verifiable rewards (GRPO). Then I put the model in the browser with ONNX on WebGPU or WASM.
Head of Engineering, BlueRickshawNew Delhi, India
I build products end to end — and train, evaluate and ship the small models inside them.
Portfolio 2026 · seven projects, one career, no filler.
Hold the button to install.
Keep scrolling — the rest of the OS is below
README.md
I'm Srinath. I build AI products end to end — the interface, the backend, the infrastructure, and the small models inside them. I trained a 2M-parameter model that turns a spoken clearance into a command, right in your browser. I built an MCP server that tests options strategies honestly, and never calls one good. Today I run engineering at BlueRickshaw, moving food from a supplier abroad to a kirana shelf.
Applications · 07
Two of my own, five as the only developer at PLXN World — and the one I run now. Each opens as its own application.
space + time
The SkyControl reel — 9.6 seconds — cut into 32 frames and stacked like pages. Scroll moves the present through it.
Products work like this for me. Every version is still in there — I just keep cutting a better slice.
Timeline
2020 – 2023
Computer applications. I was shipping apps before I graduated.
Jan 2023
The only developer. Five production products over the next three and a half years.
Jan – May 2023
Multi-tenant B2B distribution backend: FEFO stock and nearest-warehouse routing.
Jan 2023 – Mar 2024
Social dining app, shipped to the App Store and Google Play.
Jun 2023 – Mar 2024
Tutoring marketplace with a credit engine and teacher payouts.
Sep 2024 – Apr 2026
Syllabus-aware AI tutor: RAG, an agent, and three tiers of memory.
Mar – May 2026
Live meeting transcription and fact-checking, as a PWA.
May 2026
Voice ATC on small models, and an import platform, in the same month.
Aug 2026
An MCP server for options backtesting.
Oct 2026
From a supplier abroad to a kirana shelf.
13 Nov 2026
On the official MCP Registry, npm and PyPI.
14 Nov 2026
MIT licensed. Models on Hugging Face.
What I do
01
I train from scratch, fine-tune (LoRA or full) and run RL with verifiable rewards (GRPO). Then I put the model in the browser with ONNX on WebGPU or WASM.
02
RAG, agents, MCP servers, voice pipelines, document intake.
0.6%
word error rate on real ATC recordings, from a 110M speech model
New Delhi
--:--:-- IST
03
Next.js and React, FastAPI, Node, Flutter. Postgres, Redis, ClickHouse, Firestore.
04
Docker, Nginx, CI/CD with OIDC publishing, Prometheus, Grafana and Loki, security reviews.
372M
one-minute option bars behind every Stratify backtest
edu
BCA
Vellore Institute of Technology, 2020–2023
7
products shipped — five of them as a sole developer
Contact
I'm in New Delhi and open to relocation. The fastest way to reach me is email.
Aeon Labs: openatc.aeon-labs.site · stratify.aeon-labs.site