# harshal rudra
  masters @ iiitb '27 · bangalore, in

Building AI systems @outmarket.ai and researching audio-driven talking-head synthesis at the multimodal perception lab, iiit bangalore. Trying to understand how machines learn, through linear algebra, statistics & calculus. And still trying to find a match on bumble.

## currently
[outmarket.ai] ai software engineer · since may 2026
Working across the AI systems that turn unstructured information into structured, trustworthy data.
- config-driven systems replacing repetitive hand-written logic
- LLM extraction → structured, source-attributed records
- human-in-the-loop review interfaces · production hardening
// kept light, it's their code not mine
[mpl-iiitb × openstream.ai] ml research · jan–may 2026
Audio-driven talking-head synthesis, a speaking face from a voice and a single image.
- motion-first FOMM redesign: keypoints → motion → render
- diffusion stabilization under mixed-precision (AMP)
- two-stage audio→landmark→render; pivot to FLAME 3D face
## projects
███│ ▒▒▒▒▒▒▒▒▒▒▒▒
lightsout ↗ web performance, measured honestly  [npm]
Computes the earliest paint the network allows, then benchmarks it against 77 real sites: actual paint ran 10.7× later, correlation ≈ 0. Built it to prove the 14 KB rule and published the result that killed it. boxthis is the fix, a zero-dependency streaming shell that paints at the floor.
● ● ╲ ╱ ●─────●─────● ╱ ╲ ● ●
helios ↗ AI policy analysis & risk-gap engine
LLM triplet extraction builds a coverage knowledge graph (NetworkX, not vector RAG), then a two-sided matcher runs business-risk ⟷ policy-coverage gap analysis.
┌───┐ ───── ▸→ │███│ ───── │███│ ───── │ │ ───── │ │ └───┘
archimedes ↗ AI research assistant  [live]
Searches arXiv + OpenAlex, analyzes every paper with an LLM, and synthesizes the evidence into one structured, citable answer, streamed live and exportable as a PDF.
◆ · ◆ · ◆
learning navigator ↗ a course map, not a course list
Topics render as regions and resources as points of interest, so you navigate a subject instead of scrolling it. Peer recommendations from sentence-transformers + KeyBERT, plus a diffusion image studio behind a multi-GPU queue. Solo, four weeks.
┌─┬────────┐ │▪│ ▌ │ │▪│ │ │ │ │ └─┴────────┘
old-monk ↗ an AI-native terminal workspace
Written in Zig on top of Ghostty. PTY-backed sessions for coding agents that keep running and streaming while hidden, and a workspace that survives restarts. Named after the rum; a sobering amount of work went in.
┌──────┐ │░░░░░░│ │▒▒▒▒▒▒│ │ │ │ │ └──────┘
skinwise ↗ on-device diagnosis
Android app that diagnoses diseases with an LLM running fully on-device, no servers, nothing leaves your phone. Funded by the state government.
██ ██ ██ ████ ████ ██████████ ──────────∑
eclipse ↗ on-device expenses
Android app that tallies your expenses in seconds and, as a side quest, dishes out personal-finance advice, all on-device, zero privacy invaded. Funded by my university.
## writing
001  How I Built a Neural Network from Scratch Without Losing My Mind
a complex network reduces to matrix multiplications.
002  The Curious Case of Well-Behaved Matrices
orthogonal init keeps gradients from vanishing/exploding.
003  The Persistent Orthogonality of Trained Weight Matrices
weights stay orthogonal long after training ends.
## pins
a wall of photos →
random things i've pointed a camera at.
## connect
github · twitter · linkedin · instagram
open to freelance and contract work · harshaldrive3073@gmail.com
let's build something that learns.