Reyaz Khan
Opening the
Black Box.
Full-stack engineer building AI-native products. I also research mechanistic interpretability to reverse-engineer how neural networks implement algorithms internally, circuit by circuit.
System Architecture.
I'm a Rutgers CS graduate (May 2026), focused on building AI-native products and researching mechanistic interpretability.
My background spans full-stack development and ML. I've shipped agentic tools, worked across the stack from React to Python backends, and done original research on how transformers implement boolean logic internally.
Featured Engineering
Fawn: AI Daycare Receptionist
Autonomous Twilio phone agent for daycare centers. Speaks via Gemini Native Audio, schedules tours, generates PDFs, and syncs live to a dashboard.
Computer Vision Alarm Clock
A Next.js alarm clock application that uses TensorFlow.js and webcam hand-tracking to silence alarms. To turn off the sound, users must raise both hands in front of the camera.
OddsTracker: Sports Odds Aggregator
AWS serverless backend that aggregates sports odds from public APIs into DynamoDB, using API Gateway and Lambda to trigger threshold-based notification alerts.
Mechanistic Interpretability bridges the gap between deep learning outputs and human reasoning by deciphering the exact algorithms learned by neural networks.
"""Visualizes the induction heads driving context continuation."""
activations = model.run_with_cache(...)
return activations.plot()
By treating modern LLMs as a highly advanced compiled binary, we can reverse-engineer features, allowing us to edit factual knowledge predictably rather than training from scratch.
Get In Touch
Rutgers CS Graduate (May 2026). Open to SWE and AI roles.