I build computational systems where models, performance, and hardware constraints meet.
My work sits around low-level systems, AI infrastructure, high-performance computing, simulation, and quantitative software. I am usually drawn to problems that start as research questions but need to become inspectable, measurable systems.
models / representations -> systems / runtime behavior -> hardware / latency constraints
AI tooling -> C/C++ infrastructure -> GPU / FPGA / HPC
market state -> parsers / simulators -> low-latency paths
| Direction | What I care about | Typical artifacts |
|---|---|---|
| Systems & performance | memory behavior, concurrency, binary formats, runtime cost | parsers, engines, tooling, C/C++ systems |
| AI systems | retrieval, local models, evaluation, representation analysis | RAG pipelines, model probes, experiment harnesses |
| HPC / hardware-aware software | GPU computing, FPGA-adjacent design, simulation | CUDA-oriented work, VHDL/Verilog experiments, numerical code |
| Quant infrastructure | market data, order books, state modeling, latency | ITCH parsers, order book builders, backtest/simulation tools |
| Project | Technical angle | Why it matters |
|---|---|---|
| rag-project | retrieval, reranking, local LLMs, hallucination gates | makes model behavior measurable instead of demo-only |
| nasdaq-itch-parser | C++ binary protocol parsing, redundant feeds, arbitration | market data is a good testbed for correctness under throughput pressure |
| nasdaq-order-book | event-driven order book construction in C++ | turns exchange-style event streams into inspectable market state |
| market-state-representation-learning | representation learning for market regimes | explores how financial state can be modeled beyond raw time series |
| Model-Based-RL | dynamics modeling, planning, reinforcement learning | connects learned models with control and simulation |
| cloud-computing | services, queues, databases, containerized infrastructure | practical distributed-system substrate for real workloads |
- Protein and scientific foundation models: representation spaces, probing, and model-assisted discovery.
- Hardware-aware low-latency systems: market data, FPGA/GPU-adjacent computation, and performance paths.
- Systems for AI: local inference, evaluation, retrieval quality, and tools that expose model behavior.
- Build small but complete systems that can be measured.
- Prefer explicit interfaces, reproducible experiments, and inspectable failure modes.
- Treat math, model behavior, latency, and hardware constraints as parts of the same problem.