AI coding and compute
The architecture pages define concrete artifacts and gates, while shipped projects remain visibly separate.
A provider-neutral harness for Codex and Claude
One typed task protocol, isolated worktrees, normalized event traces, executable test gates, independent review, cost and latency accounting, and routing learned from matched tasks. The agent is replaceable; the durable runtime and evaluation are the product.
Portfolio roadmap CUDA + architectureFrom vector addition to a fused transformer block
The LeetGPU list reorganized as an architecture curriculum: coalescing and elementwise fusion; reductions, scans and histograms; tiled GEMM and convolution; quantization and attention; then full-block capstones. Every stage requires a correctness oracle, profiler trace, roofline position, variance, and comparison with a production baseline.
Systems with live or exact evidence
One uses exact combinatorial references; the other exposes the service-level math of a running deployment.
Chess that proves itself before it learns
The move generator matches published perft counts over more than ten million positions. A classical engine is complete; the AlphaZero-style self-play path is still training.
Live service · verified Jul 25, 2026
Reliability
Uptime and SLO monitor
A self-hosted probe service with latency and availability SLIs, rolling error budgets, and multi-window burn-rate alerts.
Kernel behavior to accelerator design
The useful portfolio bridge is a quantitative chain from algorithm to dataflow, memory traffic, RTL, and system cost.
Measure the operation
Arithmetic intensity, occupancy, divergence, locality, and numerical error.
Choose what stays put
Weight-, output-, or row-stationary schedules and explicit SRAM/HBM traffic.
Build the block
Quantized MAC array, scratchpads, DMA/AXI, control, cycle model, and assertions.
Prove the tradeoff
Timing, area, power proxy, utilization, compiler schedule, throughput, latency, and cost.
FPGA matrix-multiplication accelerator
The existing capstone for pipelining, parallel MACs, memory movement, interfaces, and verification.
Advanced systems architecture
HBM, tensor cores, systolic dataflow, precision, sparsity, interconnects, and rooflines.
Parallel computing
Measured CUDA and Triton work on a real H100, from memory behavior through high-level kernels.
Software foundations
Implementation notes behind the projects, kept one layer below the employer-facing work.
C++ field guide
Ownership, lifetime, RAII, keywords, generic programming, ranges, container internals, concurrency, and C++11 through C++26.
C++ problem-solving patterns
Seven focused chapters with recognition signals, invariants, baseline-to-optimal progressions, grouped examples, and tested C++20 implementations.
Algorithms and data structures
Canonical implementations across Python, C++, Rust, and TypeScript with tests.
Systems design
Requirements, capacity, data models, failure modes, and tradeoffs for large services.
Open-source internals
Guided code-reading paths through PyTorch, vLLM, Redis, SQLite, and other production systems.