
New Brunswick, NJ · open to SWE / DS / AI
PULKITCHAUDHARY
Quadruple major · Computer Science · Data Science · Mathematics · Statistics
AI engineer at a YC startup, where I build multi-agent pipelines on MCP and LLM tool-use that research, write and deploy production software end to end. ML researcher at WINLAB. I teach data science at Rutgers, keep a hand in quantitative research, and spend my weekends at the kitchen line.
- Agents
- Multi-agent pipelines on MCP and LLM tool-use, shipping real software end to end
- Retrieval
- RAG over real corpora, so an assistant answers from sources instead of guessing
- ML research
- How small a network can be and still solve the task, benchmarked honestly
WINLAB- School
- Rutgers · New Brunswick
- Quadruple major
- Computer Science · Data Science · Mathematics · Statistics
- GPA
- 3.85
- On court
- AI Eng @ Blaze (YC S24)
The Season
The roles I'm playing now, in order of the clock. Follow the ball.
- AGENTIC
Blaze
AI Eng · Y Combinator S24 · May 2026 — PresentEnd-to-end agentic pipeline: raw lead → live production site.
TEAM 5-personPIPELINE Research → Implementation → ReviewSTACK MCP · Azure AI FoundryVisit Blaze →
RESEARCHWINLAB
ML Research · May 2026 — PresentA minimal CNN policy that solves mazes one move at a time.
BENCH 1,000 mazesARCH 5 conv + dense headINPUT 6-channel gridVisit WINLAB →
TEACHINGData 101 Lecturer
Rutgers · PresentRun weekly R recitations + stats labs; grade 250+/mo on time.
STUDENTS 40+GRADED 250+ / moON-TIME 100%Rutgers CS →
QUANTQuant Finance
Rutgers QFC · Feb 2026 — PresentCalibrate SABR stochastic-vol on real options via least-squares.
ERROR < 2 bpsCHAINS 500+MODEL SABRVisit Rutgers QFC →
Skill Rally
Strike a ball to pull the production log behind it.
The Work
Five shipped builds, each with its own film room: architecture, metrics and the code. Click any card.
The Lab
ScarletAI, the maze policy network and the AI-SDE engine are not screenshots here. These are the actual algorithms, running right now in your browser.
Maze Policy Network
winlab · conv policyready · maze 1
The environment my research runs on, live: a procedurally generated maze, BFS expanding to the optimal path (the ground truth I score against), then an agent stepping it one move at a time. The trained policy itself is a 5-layer conv net in PyTorch, benchmarked over 1,000 of these.
AI-SDE Engine
quant · monte carloready
Real geometric Brownian motion, simulated in your browser: 220 price paths under 32% vol, and the terminal distribution they build on the right. My production engine swaps the constant vol for a learned model in PyTorch and runs 2,000+ paths per second.
ScarletAI
rutgers agent · raganswer, grounded inCampuses
Rutgers New Brunswick is spread across several campuses including College Avenue, Busch, Livingston, Cook and Douglass, which is why getting between classes usually means a bus.
live tf-idf cosine over a labelled sample corpus · production: full Rutgers corpus, vector store, routed across Gemini · Groq · Ollama · Anthropic
no video, no gifs: every frame is computed live
Trophy Case
Hardware and the coursework behind it. Push the shelf.
Your serve.
Hiring, building something hard, or down for a game? Put a serve on the line.