A mix of work, research, and side quests.
Relling — Robot Deployment Systems
As a founding engineer at Relling, I'm standing up a bimanual UR5e work cell end to end: a Go control plane that owns robot drivers, camera health, and cell telemetry; the calibration and perception stack; a scripted manipulation baseline instrumented with a typed failure taxonomy and per-trial condition snapshots; and the data engine that takes teleop episodes through quality gates into VLA fine-tuning. Currently evaluating a fine-tuned open-weights VLA against the scripted baseline, with per-layer failure attribution. Active R&D — deployment in progress.
GELLO Teleop Rig + Data Engine
A GELLO-style teleoperation rig for the cell, plus the episode pipeline behind it: MCAP recording, automated quality gates, a review dashboard, and conversion to LeRobot format for policy training.
Distillery
Egocentric-video annotation at cloud scale — the distributed pipeline we ran before pivoting to robot deployments.
Javelin
Helped build the intelligence layer of athlete marketing at Javelin, an AI-native operating system for the athlete economy — semantic search on pgvector, a Go-based data pipeline with SQS, and an AI command center with structured-output SSE streaming. Built with NextJS, React Native, and AWS (Lambda, Amplify, Athena, Bedrock).
Earlier Work
- SinkyViT — a ViT with Sparse Sinkhorn Attention, ~15% FLOP reduction on MNIST and CIFAR-10
- FastConv — concurrent Go convolution engine and ray tracer with dynamic work stealing, up to 6x speedups
- vb-FaaS — a serverless FaaS platform on Redis, RabbitMQ, and FastAPI
- CTA Delay Predictor — real-time train-delay prediction on Kafka and Spark
- Tech Titans — an eBay clone on microservices with RabbitMQ, PostgreSQL, and MongoDB
- GAS — genomics annotation SaaS on AWS with autoscaling and free/premium tiers
- Phil — a course-catalog search chatbot for UChicago undergrads, built before GPT-3 was out
Tale of Two Cities
A 25-page mini-thesis comparing Indian immigrant experiences in Doha and San Francisco, examining how educational level shaped socio-economic outcomes. Methodology included scraping geo-tagged tweets for sentiment analysis.
Where's My Uber
A shorter paper on the formation of social microstructures by Uber drivers in response to the information asymmetry of the rideshare market.
Lumiere — Beli meets Letterboxd*
A rank-based movie aggregator built in Swift, combining the UX of our generation's two most beloved apps (citation needed).
Other Projects
- GMM and BMM models with numpy
- GANs and VAEs in PyTorch on CIFAR-10 and MNIST
- SQL Query Engine for Course Catalog
- Movie about the invention of a Circle
- Breaking Bad Cinematography Recreation
- Scheduler for IB Students in Flutter
- SASA 2024 Trailer