B.E. CSEEst. 2029CGPA 8.03

01 — Intro

Sri Ramaanuja Dasan

Sri Ramaanuja Dasan speaking at a podium
Presenting · 2026

AI/ML ENGINEER

I build and ship AI systems end to end — from vision-language document accessibility used daily by real teams, to blind radio-signal decoding, to bit-exact model quantization merged into llama.cpp.

Selected work↓

02 — Selected Work

Things I built and shipped.

— Playground

Poke around in
my shell.

visitor@srd — zsh

try help, ls, open rawrs, neofetch  ·  ↑/↓ history · Tab completes

how_i_work.py
▋

03 — About

A second-year CSE student who ships like it's the job.

I'm building AI that reaches real users — a tool used daily at WinVinaya, a hackathon-grade audit engine, a signal decoder from first principles. I test design decisions against benchmarks and held-out sets, and I care about systems that stay reviewable and reversible. Long game: found my own company.

Winner ×2SJIT internal Smart India Hackathon selection — 2025, 2026
3rd · QuantathonSRM University
3rd · Cyberfest '26KCG College of Technology
Special MentionV-VORTEX, VIT Chennai · AgriSense at E-Summit, IIT Madras

— What I do

How I'm useful.

def apply_ml()01

Applied ML & LLMs

Vision-language pipelines, sentence embeddings, anomaly detection and survival models — measured against benchmarks and held-out sets, never vibes.

return RAWRS — shipped to WinVinaya, daily
def ship_end_to_end()02

End-to-end delivery

Data pipeline → API → UI, by one person: FastAPI + Pydantic on the back, React + TypeScript on the front, Docker and CI around it.

return idea → production, one hand on it
def build_from_scratch()03

First-principles systems

When there's no library for it, I build the hard part myself — Viterbi decoders, bit-exact quantization, statistical significance tests.

return ICHNOVA — 30/30, 0 false accepts
def contribute()04

Open-source infra

I contribute to the tools I use. I read code I didn't write and leave it better than I found it.

return llama.cpp #30630 — merged

— the toolbox

What I build with.

AI / ML

  • Qwen2.5-VL
  • Sentence Transformers
  • scikit-learn
  • GGUF quantization
  • Survival analysis
  • Viterbi / cumulants

Languages & frameworks

  • Python
  • TypeScript
  • React
  • Next.js
  • FastAPI
  • Pydantic
  • Tailwind

Data & documents

  • NumPy
  • SciPy
  • pandas
  • Docling
  • PyMuPDF
  • Surya OCR
  • python-docx
  • WCAG 2.x

Tooling

  • Git
  • Docker
  • CI
  • pytest
  • Vercel
  • Firebase

— the path so far

How I got here.

  1. Jun 2026 — Present

    Forward Deployed Engineer — AI Document Intelligence & Accessibility

    WinVinaya InfoSystems · Bangalore

    Sole developer of RAWRS, a Qwen2.5-VL document-accessibility platform remediators use almost daily — ~3h per document down to 10–20 min. Cut reviewer workload ~70% by finding and merging ~385 duplicate checks, no evidence dropped.

  2. Oct 2026

    Open-source contributor — llama.cpp

    ggml-org · PR #30630, merged

    Added NumPy Q1_0 & Q2_0 GGUF de/quantization to gguf-py so Python tools can load 1- and 2-bit models. Bit-exact with the C reference — 66/66 checks across 27 quant types.

  3. 2026 · SIH

    Thadam — team lead of 6 · ICHNOVA — sole developer

    Smart India Hackathon 2026

    Thadam: audit intelligence over 476,781 lifecycle events — sentence embeddings, learned archetypes, Cox survival risk. ICHNOVA: a blind radio-signal decoder from first principles — 30/30 held-out, 0 false accepts, verified on real WWV / WWVB / DCF77 signals.

  4. 2025 — 2029

    B.E. Computer Science & Engineering

    St. Joseph's Institute of Technology, Chennai · CGPA 8.03

    2× winner of SJIT's internal Smart India Hackathon selection (2025, 2026). 3rd — Quantathon (SRM) · 3rd — Cyberfest '26 (KCG) · Special Mention — V-VORTEX (VIT Chennai).

Full résumé — PDF↗

— the one you're looking at

Built to ship.

Second year of CSE, already in production a real company runs on. I decode radio from nothing but the raw signal, and I merged into llama.cpp while most people are still fixing their first bug. No demos, no maybes — I build it, then make it prove itself. Move across the frame; it sharpens where you look.

04 — Open Source

Contributing to the
infra I use.

ggml-org / llama.cpp ● Merged · PR #30630

NumPy Q1_0 & Q2_0 quantization for gguf-py

Added 1-bit and 2-bit GGUF de/quantization to the Python library so tools can load models (e.g. Bonsai) that previously failed with NotImplementedError. Matched the C reference bit for bit — 66/66 checks across 27 quant types.

Read the PR ↗

— live from GitHub

Receipts, fetched live.

—Public repos
—Languages shipped
—Last pushed
—On GitHub since

    fetching @SomeNobody21112 …

    — the merge, running

    Float32, folding
    to two bits.

    The quantizer I shipped to llama.cpp, made visible: a plane of 32-bit weights collapsing block by block onto just four levels — the whole trick behind Q2_0. A model gets roughly sixteen times lighter, and still speaks.

    3hRemediation time cut with RAWRS
    0Tests gating every RAWRS change
    0Lifecycle events analysed in Thadam
    0Signals decoded, 0 false accepts

    — always cooking

    The next one's loading…

    move your cursor through it

    — you reached the end, comrade

    Meet Vladimir.

    Vladimir “Longcat” Pawtin — a communist black cat, allegedly both alive and dead. Pet him, feed him, point the laser, or ask him anything about me.

    Affection

    05 — Contact

    Let's make something real.

    Email↗ GitHub↗ LinkedIn↗ Résumé↗

    sriramaanujadasan2007@gmail.com · +91 73052 89070