> hello
I like taking systems from first sketch to working production shape: benchmarked vLLM serving on H100s, implemented distributed transformer parallelism, and built clinical ML pipelines over messy real-world data.
63 req/s
multimodal inference serving
8 GPUs
distributed training benchmarks
110M+
ICU patient-hour records modeled
Applied ML systems, distributed training, and backend projects with measurable outcomes.
featured system
featured system
featured system
2.10M tokens/sec on 8-GPU data parallel training runs
Built the parallel training stack for a GPT-style model as a capstone for a hands-on distributed systems workshop, implementing data, tensor, pipeline, context, and expert parallelism from scratch, then benchmarked throughput, memory, and convergence across 1–8 GPU configurations.
63.05 req/s peak text throughput; 916 ms p95 TTFT on mixed multimodal traffic
Benchmarked Qwen3-VL-4B-Instruct on Modal with vLLM, comparing H100 replica scaling, tensor parallelism, concurrency limits, and mixed multimodal traffic to find a low-latency, high-throughput serving configuration.
110M+ hourly ICU records; 0.995 AUROC deterioration prediction
Built an ICU early deterioration warning system on 110M+ hourly records from 50,920 patients (MIMIC-IV), using XGBoost with 12-hour rolling window features to predict vasopressor initiation, intubation, CRRT, or death within 12 hours