GPU Optimization · LLM Training & Serving · Distributed Systems · Backend Engineering
I build and evaluate machine-learning systems from model training through inference, performance analysis, and deployment. My work spans LLM fine-tuning and retrieval, GPU kernels and compiler optimization, distributed training, and fault-tolerant backend systems.
Jun 2025 – Present
- Engineer production ML/LLM systems across data processing, retrieval, ranking, evaluation, and inference services using Python and PyTorch.
- Internal evaluation work reported a 21% retrieval-relevance improvement and 30% fewer hallucinations; model/task tuning improved measured accuracy from 78% to 86%, while a serving optimization reduced p95 latency from 1.2 s to 420 ms.
- Support deployment validation, model/version management, monitoring, and reliable inference operation.
Aug 2024 – May 2025 · Arlington, Texas
- Built TopGPT, fine-tuned on 3+ textbooks and paired with retrieval-augmented generation over 1,000+ research-paper chunks stored in Pinecone.
- Researched CTMap/DTMAP, combining LLM-enabled path planning, Dijkstra-generated routes, OpenStreetMap, digital twins, and NVIDIA Sionna for connectivity-aware mmWave/6G mobility.
- Supported Numerical Methods instruction for 50+ students using Python and MATLAB.
Jun 2019 – May 2023
- Built and operated Java/Spring Boot services across development, staging, and production, scaling the service footprint from 20 to 50+ services.
- Resolved 20+ critical incidents and improved reported SLA compliance from 95% to 99.9%; backend capacity increased 40% while infrastructure overhead fell 20% through service, database, and connection-pool optimization.
- Conducted 100+ code reviews and mentored 5+ engineers.
| Project | What it demonstrates | Reported benchmark / evidence |
|---|---|---|
| metal-inference-server | Continuous-batching LLM inference for Apple Silicon, paged KV cache, streaming, cancellation, metrics, and traces | About 1.5× throughput improvement from continuous batching in the documented M-series Mac setup; negative quantization results are also documented |
| LLMBoost | MLIR compiler pass for RMSNorm → Linear fusion | 1.63× FP16 speedup for the documented [512, 4096] × [4096, 4096] shape (calculated from 0.340 ms / 0.208 ms); a separate dtype summary reports 1.55× BF16 |
| flash-attn-from-scratch | Triton FlashAttention, int8 GEMM with fused dequantization, and fused CUDA kernels | NVIDIA A30 measurements with per-sequence-length comparisons against PyTorch SDPA |
| distributed-training-models | PyTorch DDP/FSDP, mixed precision, and pipeline-parallel training | 3.14× AMP throughput versus FP32 in the reported single-A30 setup; FSDP memory results documented by configuration |
| search-engine-rs | Rust tokenizer, inverted index, BM25, sharded query execution, and HTTP control plane | Median 1.55× postings-intersection speedup and 4.35× batch-query speedup across five runs on 82,360 MS MARCO passages |
| ledgerdb | Go distributed commit log/document store with Raft, MVCC, WAL, TCP transport, MinIO, and fault injection | Benchmark artifacts and network-failure scenarios are committed with the project |
More projects: all public repositories.
- DTMAP — Digital Twin-Guided AI Path Planning for Connectivity-Aware Mobility. IEEE Open Journal of the Communications Society, 2026. DOI
- CTMap — LLM-Enabled Connectivity-Aware Path Planning for mmWave 6G Networks. IEEE ICC 2026. arXiv preprint
- DT-LAWN: Incrementally Synchronized Digital Twins for Real-Time NextG Wireless Navigation. ACM, 2026. ACM Digital Library
- Efficiency–Equity Trade-Off Characterization in Hierarchical Traffic Congestion Pricing. IEEE ITSC 2026. Paper PDF
- ML / LLM: PyTorch, JAX, Hugging Face Transformers, RAG, SFT, LoRA/QLoRA, model evaluation, embeddings, reranking
- Inference / GPU: CUDA C++, Triton, FlashAttention, vLLM, KV cache, quantization, kernel fusion, ONNX Runtime, TensorRT
- Distributed systems: Go, Java, Rust, Raft, replication, MVCC, WAL, sharding, fault injection, microservices
- Infrastructure: AWS, Docker, Kubernetes, CI/CD, Prometheus, OpenTelemetry, PostgreSQL, Redis