Vignesh K.
Bengaluru, India · Building at Streamoid

I turn hard ideas into
products that ship.

Solutions architect and hands-on engineer with 14 years of experience turning difficult ideas into dependable products across AI agents, real-time 3D, cloud systems, and developer tools.

14years building
production systems
01published
US patent
70%faster catalog
turnaround
3+enterprise brands
in production

Intellectual property

✦

Published US patent

Automated Product Video Generation for Fashion Items

US 2024/0078576 A1 ↗
View on
Google Patents ↗

Selected work

Systems with consequence.

Flagship platforms where architecture, product thinking, and execution had to work as one.

01 / 03

AI-native design platform

Artifax

An infinite, GPU-accelerated design canvas orchestrated by planner–executor–validator agents, built for enterprise creative workflows.

↳Used by Puma, Ajio & ABFRL

  • LangGraph
  • WebGL / WASM
  • Multi-agent AI
View platform ↗
Artifax product interface
02 / 03

Real-time 3D platform

Desara

A cloud-streamed Unreal Engine studio that brings photoreal 3D product design to the browser, without workstation-class hardware.

↳Adopted by London College of Fashion

  • Unreal Engine
  • WebRTC
  • GPU orchestration
View platform ↗
Desara product interface
03 / 03

Enterprise ML systems

Catalogix

Automated cataloging and AI photoshoot infrastructure built around distributed classifiers, diffusion models, and custom LoRA adapters.

↳70% faster catalog turnaround

  • Computer vision
  • Diffusion
  • Distributed systems
View platform ↗
Catalogix product interface

In the lab · Currently building

VMotionX

Type a prompt and author 3D motion for characters and robots, fully offline. Pose it, extend the motion, and export straight to your DCC or rig. Real-time and GPU-accelerated.

  • Text → Motion
  • Characters + robots
  • Real-time 3D
  • On-device
Explore the project ↗

Applied research · Fine-tuning

Fifty-one photos,
one small adapter.

Adapting a 12B-parameter image model to a subject it has never seen, on hardware that fits under a desk.

The dataset · 51 phone photos, shot at home. Front, three-quarter, both profiles, the back of a head. No studio, no lighting rig. Shown here downscaled.
Generated image: black suit beside a helicopter Generated portrait in front of the Eiffel Tower Generated side profile in front of the Eiffel Tower
Generated from prompts. None of these places, clothes or framings exist anywhere in the dataset.

LoRA

Full fine-tuning rewrites every weight in the model, which is why it costs what it costs. LoRA freezes the base model and trains a pair of small low-rank matrices alongside the attention layers instead. What ships at the end is an adapter measured in megabytes. The base weights stay untouched, so one model serves every subject you train.

QLoRA

QLoRA shrinks the frozen base further by quantizing it to 4-bit, then trains the adapter on top of it in higher precision. The memory saved is the whole point. It moves a run that would have demanded datacenter cards onto a single consumer GPU, which is what makes an experiment like this something you can start on a weekday evening.

Recognition

“These students are being given the opportunity to trial an exciting new AI research and product development tool, Desara AI.”

Julia Redman · Lecturer & Unit Lead, London College of Fashion

LinkedIn post about Desara AI at London College of FashionView original post ↗

In the press

Paparazzi Pass Lets Ski Resort Visitors Take Slope-Side Selfies

RFID Journal

Paparazzi Pass Lets Ski Resort Visitors Take Slope-Side Selfies

BLE beacons and an app from DejaView Concepts automatically capture action videos of skiers and snowboarders as they pass through beacon zones, letting them view, edit, and share the footage.

Read on RFID Journal ↗

Certifications

Credentials.

View resume →

How I work

Architecture is a product discipline.

01

Start with the constraint

The best architecture emerges from a precise understanding of the user, the economics, and the operational reality.

02

Prototype the hard part

I de-risk the unknowns early (rendering, latency, model behaviour, and scale) before the surrounding system becomes expensive.

03

Stay close to the code

Strategy is strongest when grounded in implementation. I lead architecture while remaining hands-on with the critical path.

Writing & explorations

Notes from building.

View Medium profile ↗

Medium · Technical writing

Offline Data Transfer on Android Using Ultra Sound FSK & FFT

Exploring robust offline data transfer using ultrasonic audio, frequency-shift keying, and signal processing on Android.

Read article ↗

Medium · Technical writing

A Decentralized Ride-Sharing DApp on Ethereum

How a weekend hackathon project explored ride-sharing infrastructure using blockchain and IoT.

Read article ↗

Medium · Technical writing

When This OMR Scanner Was a Computer Vision Feat

A look back at building a smartphone OMR scanner when mobile computer vision was still genuinely difficult.

Read article ↗

Always exploring challenging problems

Working on something
hard and interesting

vignesh0702@gmail.com ↗