Autonomy
End-to-end self-driving stack
A nine-layer control stack, from camera capture and TensorRT inference down through firmware to CAN, that steers and brakes a production car on public roads.
2023 Mazda CX-5 · Jetson Orin Nano · DIY STM32F407 panda
- Lateral loop at 50–70 Hz, steering torque on CAN 0x243
- Bare-metal STM32F407 gate: 1.5 Mbaud UART, 2,325 CAN frames/s
- Traced a 99.8% packet-loss stall to a 16-instruction USB race
Systems
Repo RFQ engine
A request-for-quote lifecycle service for securities financing. Two dealers can never double-book the same quote: row locks, state re-validation on accept, and a unique constraint as the backstop.
- 20 concurrent goroutines → exactly 1 trade, 19 clean rejections
- int64 minor units and integer basis points — no float drift
- 91% coverage on an I/O-free domain layer
Low-level
Filesystem & PCIe transaction layer
A disk-backed Unix filesystem in C — i-node allocation over a bitmap, direct and single-indirect addressing, full path traversal — alongside a bit-level PCIe transaction-layer packet parser in C++.
- Complete file and directory API, verified by 50+ tests
- Byte-enable masking and tag-tracked outstanding reads
- Zero-copy over the caller's buffer: 10K packets in ~10 ms
Networking
Protocol design & concurrent servers
AFLENT, a transfer protocol built around a packed three-byte header — 6-bit array ID, 5-bit fragment, 10-bit length, three flags — parsed by bitmasking and reassembled correctly out of order.
- 64 concurrent transfers from out-of-order delivery
- Block cipher expanding a 64-bit key to 1024 bits of round state
- Texas Hold'em server multiplexing six players, non-blocking
Capabilities
What I work in
Grouped by where it actually gets used, not ranked by a number I made up about myself.
Languages
- C
- C++
- Rust
- Python
- Java
- CUDA
- Go
- TypeScript
- SQL
- ARM / MIPS Assembly
Embedded & automotive
- STM32F4 bare-metal firmware
- CAN / CAN FD
- UDS diagnostics
- DBC
- DMA
- UART / SPI / USB
- Interrupts & timers
- SWD debugging
- Hardware-in-the-loop testing
- panda safety models
Systems & AI/ML
- Linux systems programming
- Real-time scheduling
- Concurrency
- POSIX sockets
- PCIe
- Protocol design
- PyTorch
- TensorRT
- tinygrad
- Jetson Orin
- Camera calibration
- Sensor fusion
Web & cloud
- React
- Next.js
- Node.js
- AWS (Lambda, DynamoDB, S3)
- Docker
- CI/CD
Index
Everything else
Autonomy & AI
- End-To-End-Self-Driving
comma.ai's openpilot model on a Jetson Orin Nano driving a 2023 Mazda CX-5 through a DIY STM32F407 panda. Standalone monorepo.
CC++CUDA - cuda-kernels-from-scratch
Hand-written CUDA kernels built to learn GPU parallel-programming fundamentals from first principles.
CUDAC++GPU - disconnect-four-solver
Connect-Four engine with a search-based AI opponent in C++.
C++AlgorithmsGame AI
Systems & Low-Level
- repo-rfq-engine
RFQ lifecycle engine for repo in Go and MySQL with an Angular blotter. Exactly-once execution under concurrent acceptance.
GoMySQLAngular - linux-filesystem
Disk-backed Unix filesystem: i-node allocation over a bitmap, direct and single-indirect addressing, full path traversal.
COperating Systems - pcie-tlp-protocol
Bit-level PCIe Transaction Layer Packet parser with byte-enable masking and tag-based tracking of outstanding reads.
C++Hardware Protocols - AFLENT-Network-Protocol-Implementation
Transfer protocol with a packed 3-byte header, reassembling 64 concurrent transfers from out-of-order delivery.
CSocketsProtocol Design - C-Multiplayer-Poker-Server
Texas Hold'em server multiplexing six players over POSIX sockets without blocking.
CNetworkingConcurrency - MIPS-Red-Black-Tree
Red-black tree implemented directly in MIPS assembly, including rebalancing.
AssemblyData Structures