projects
Selected work in agent systems, LLM inference, model serving, retrieval, and applied machine learning.
Agentic AI
Human-in-the-Loop Agent Systems
Agent workflows that combine structured planning, MCP tool access, traceable execution, and explicit human approval.
LangGraph, MCP, FastAPI, OpenTelemetry
View detailsLLM Inference
FlashAttention V2 from Scratch
CUDA implementation of FlashAttention with tiled attention, online softmax, and Nsight-guided optimization.
CUDA C++, PyTorch, Nsight Compute
View detailsModel Serving
QuantRoute
Quantization-aware inference router that trades precision, latency, and cost across multi-precision serving lanes.
vLLM, llm-compressor, AWQ, GPTQ, NVIDIA A100
View detailsLLM Inference
Speculative Decoding from Scratch
A complete draft and verify implementation with correctness derivations and performance benchmarks.
PyTorch, Hugging Face Transformers
View detailsRetrieval and MLOps
MultiDocChat
Production retrieval-augmented generation service with tracing and GitOps deployment.
FastAPI, LangChain, FAISS, Kubernetes, ArgoCD
View detailsModel Serving
Continuous Batching Inference Server
A minimal serving engine that demonstrates the scheduling ideas behind high throughput LLM inference.
Python, PyTorch, LLM Inference
View detailsApplied Machine Learning
Probabilistic Yard Gain Forecasting
Calibrated play-outcome modeling from NFL tracking data, framed as a full cumulative distribution rather than a point estimate.
Python, Scikit-learn, XGBoost, Pandas, NumPy, Flask
View details