Software engineer at HUB9. I build AI products, and the engines under them.
Python, FastAPI, React and LLMs by day. On my own: an inference engine for Apple Silicon, a robot arm that plans from one sentence, a ride app for Indore and a language app for kids.
|
SmolVLA goes from three cameras to 50 robot actions in 267 ms, 2.3x faster than PyTorch on the same Mac. LLM decoding is level with llama.cpp itself. C++ Metal Core ML llama.cpp Python
Website · Source (Apache-2.0) |
94.7% accuracy over 150 trials with Claude. TypeScript MuJoCo WASM Claude Three.js Python
Try it in the browser · Source (Apache-2.0) |
FastAPI PostGIS Redis Flutter Next.js
Website |
Expo FastAPI LLMs Text to speech
Website · Instagram |
Software Engineer since September 2024, remote for San Francisco, working across four products.
Motion Intelligence · a platform for packaging, deploying and running custom video models
- Lead engineer on the Python CLI that packages models into Docker images
- Built the official Node.js and TypeScript SDK, cutting client integration effort by 40%
- Quantization and GPU tuning cut inference latency by 45%; designed a streaming 3D-reconstruction pipeline on S3
DBX Studio · an AI SQL assistant
- Founding engineer: architected the platform from scratch
- FastAPI backend with LLM-powered query generation across PostgreSQL, MySQL and SQLite; 15+ active beta users
SumrLabs · clinical summaries and medical-code ranking
- HIPAA-conscious FastAPI microservices; PostgreSQL schemas that cut data-retrieval time by 24%
- AWS Polly and Transcribe integration cut documentation time by 20%
June Kids · a storytelling app with personalised stories and voice narration
- Lead engineer on the React Native (Expo) app, with FastAPI services on AWS Lambda for story generation
- RevenueCat subscriptions with credit-based billing, Cognito and Google Sign-In
Before HUB9: Web Developer Intern at DevFolks, Indore (2023).
AI: LLMs, LangChain, LangGraph, RAG, agents, Claude on Bedrock
Physical AI and robotics: VLA models (SmolVLA), vision-language models, MuJoCo, NVIDIA Isaac Sim, robot-arm control from language and depth
Inference engines: dooms.cpp (my own), llama.cpp and GGUF, Metal, Core ML, model quantization
Desktop: Electron · Also shipped with: GPU optimization, video inference, 3D reconstruction, WebRTC, FHIR, RevenueCat
Certified: Oracle Cloud Infrastructure 2025: Generative AI Professional · Developer Professional · Data Science Professional · DevOps Professional
Education: B.Tech in Information Technology, IIST Indore (CGPA 8.55)
The fastest way to reach me is email: jayesh.work011@gmail.com. Based in Indore, India, and open to new roles and collaborations.