Vandit Vasa

AI engineer, currently deep in agents, voice, and evals.

Contact

San Jose, CA
Email: vasa.v@northeastern.edu
Github | LinkedIn | X / Twitter

About

I'm an AI engineer who builds like a founder. As founding engineer I took a platform to $100K+ MRR and built an autonomous software factory that runs it on its own: agents take their own tickets, ship their own code, and fix their own bugs in production. For Omi (open-source AI, 300K+ users) I rewrote the voice architecture from scratch and cut latency 10x in five days. Before that, the multi-agent pipeline behind a $250K seed, idea to production in days. Now: MS in AI at Northeastern.

If you're building something ambitious with AI, reach out.

Experience

• PracticeNova — Software Engineer Intern, Palo Alto, CA (Jun 2026 – Sep 2026)
  Built an autonomous agent workforce for dental practices; fleet technical lead running 50+ production agents completing 10,000+ client tasks/week across 300+ practices at a 95%+ eval pass rate.
• Omi (13.4k stars) — Software Engineer Intern (Contract), New York, NY (May 2026 – Jun 2026)
  Cut voice latency 10x in five days by rearchitecting the core voice spine 300K MAU depend on, unifying GPT Realtime and Gemini Live in one orchestration layer.
  Github
• Shrimad Rajchandra Aatma Tatva Research Centre — Founding Engineer, Mumbai, IN (Jun 2024 – Sep 2025)
  Sole engineer for a production platform (React Native app, Node.js/Express API, admin dashboard) serving 5,000+ monthly users, $100K+ MRR. Built an autonomous software factory where agents debug through an authenticated MCP observability layer and ship their own fixes. Grew from solo to leading a 5-person team.
  Play Store | App Store | Web | Github
• Juggernot.ai — AI Engineer Intern, Pune, IN (Sep 2023 – Dec 2023)
  Sole engineer on Marketing Swarm, a multi-agent pipeline that cut campaign turnaround from 5 days to minutes; demoed live to investors, helping land a $250K Antler seed round.

Projects

• 90 Day Agent — open-source benchmark grading frontier LLM agents on operating a simulated small business for 90 days, with 2,000+ graded tasks and LLM-judged rubrics validated at 94% agreement against human labels.
• Lux — building now: an on-device personal AI for Android on a memory engine of my own, fusing device signals into a private intent model with local LLM inference (Gemma via LiteRT), nothing leaving the phone.
• Neural Graph Drawer — a PPO + two-layer GCN layout optimizer that cuts edge crossings 51.1% vs GraphViz Neato across 11,481 Rome graphs (99.5% of the dataset), evaluated on a 32-shard SLURM run.
  code
• Cross-modal RAG — a retrieval system answering questions across 2,000+ spiritual discourse videos and 100+ books from one unified index, in English, Romanized, or native Indic script.
  code
• Edge Autonomous Navigation — a research rover that drives itself through dense crop rows: fused RealSense depth with LiDAR, ran real-time detection and avoidance on a Jetson Nano, planned paths with RRT, validated in real sugarcane fields.
• Battleship Agent — a competitive Battleship agent using probability-density targeting. Ranked #4, zero disqualifications across a 1,000-game soak test.
  code