From 0 to 1.
From 1 to n.
I've built trading engines, backend platforms, data systems, and blockchain infrastructure. I like owning a product from its first implementation through the hard engineering that follows.
Most of my work has been behind the scenes: the APIs, execution engines, databases, and infrastructure that make a product work.
I've built APIs, query engines, and transaction systems from scratch. Concurrency, caching, databases, cloud infrastructure, and the occasional production bug that takes an entire day to find.
I spent a good part of my engineering career on DeFi infrastructure. Not just interacting with smart contracts, but building the systems that find routes, calculate prices, execute trades, and talk to liquidity pools.
I've worked on machine learning and computer vision, from experimenting with model architectures to implementing inference systems. I also have experience with academic research.
A SQL engine that queries Parquet files directly from cloud storage, without importing the data into a database first.
Scanned just 0.2% of the data using three layers of query optimization. Same answer, 99.8% less data scanned.
A deployment platform that automatically isolates each customer's data, tests for security gaps, and controls what AI agents can access.
200+ simulated attacks before every deploy. Even a forgotten database filter can't expose another customer's data.
A system for rolling out features, running A/B experiments, and changing application behavior without redeploying code.
Sub-millisecond decisions, zero server calls per evaluation, and feature updates delivered in under a second.
A research pipeline that measures how prices move against market makers after their orders get filled.
From raw exchange trades to a reproducible market-risk study, across six time horizons from 1 to 300 seconds.
Matches internal trades against actual exchange fills across Binance, Coinbase, and Kraken to detect missing, partial, or suspicious executions.
Database-enforced duplicate protection, replayable exchange events, and live discrepancy detection across three exchanges.
A real-time multiplayer game backend that handles matchmaking, move validation, disconnects, rematches, and leaderboards.
The server owns every move. Clients can't cheat by changing their local game state.
A Solidity system that reads liquidity distribution across concentrated-liquidity pools, including Uniswap V4-style protocols.
Gas-bounded liquidity scanning using bitmaps, bitwise math, and low-level Solidity assembly.
Moves tokens between two blockchains by locking tokens on one chain and issuing their equivalent on another.
Validator-signed transfers with nonce-based replay protection, verified across the complete two-chain lifecycle.
Verifies blockchain transactions using cryptographic proofs and monitors on-chain activity through live event streams.
Merkle proofs, multi-signature verification, and batched on-chain checks in one verification pipeline.
An AI system that detects fires in video and predicts how they may spread using neural networks and heat-transfer physics.
92% fewer false positives, 89% accuracy on high-risk objects, running at 12–15 FPS.
A deep-learning system that analyzes endoscopic images to identify and classify ten different pathological conditions.
10 disease classes. 92.3% classification accuracy using an ensemble of specialized neural networks.
A computer vision system that tracks vehicles, detects brake lights, and identifies unusual driving behavior from video.
Three detection models working together at 30 FPS on 1080p video.
Built and operated real-time trade routing, settlement, pricing, and indexing infrastructure.
Owned async backend services, yield aggregation, data pipelines, and production operations.
Firefighters move through smoke-filled rooms with almost no information about what is about to catch next. Pure vision models are no help there: they see objects, not thermodynamics. This paper puts the physics back in with a framework that fuses real-time object detection with heat-transfer modelling, so the system reasons about conduction, convection, and radiation rather than pattern-matching flames.
Detected objects are scored for ignition risk from their material properties and thermal response, and a physics-informed neural network predicts how the fire will spread across the scene. Grounding the predictions in conservation laws cut false positives by 92% and reached 89% accuracy on high-risk object identification and spread forecasting at 12–15 FPS, fast enough to run on the equipment a crew actually carries.
Read the paper on IEEE Xplore ↗Open to backend, infrastructure, and founding engineering roles. Happy to talk systems, ideas, or opportunities.