CUDA Kernels Suite
A from-scratch suite of GPU kernels (vector add, tiled matrix multiplication, reduction, softmax, and attention) built to understand GPU performance from the hardware up. Every kernel is profiled with Nsight Systems/Compute and benchmarked head-to-head against cuBLAS and PyTorch using CUDA-event timing with warmup and median-of-many-iterations methodology.
- Hand-written CUDA kernels with standalone
nvccbuilds and self-tests - Benchmark harness plotting GFLOPS vs problem size against cuBLAS/PyTorch baselines
- Profiling-driven optimization: memory coalescing, shared-memory tiling, occupancy analysis
- In progress: full benchmark comparison table and per-kernel optimization writeups; next up: FlashAttention-style fused attention and KV caching
Production Lifecycle Management for LLM Classifiers
An MLOps framework for keeping LLM classifiers reliable after deployment, treating the model not as a one-off artifact but as a system that must be monitored, evaluated, and retrained as the world drifts away from its training data.
- Production monitoring and data drift detection using statistical tests
- Active learning strategies to select the highest-value samples for relabeling
- Incremental fine-tuning pipelines for continuous model updates
RAGstoRiches: Agentic Financial RAG System
An agentic RAG system for financial data that lets an LLM agent combine semantic retrieval with structured SQL analysis: SQL exposed as a FastMCP tool server, Qdrant Cloud for vector search, Supabase for authentication, and Databricks-hosted Postgres.
- Agentic tool use: SQL served to the agent as a FastMCP tool server
- Qdrant Cloud vector database for semantic retrieval
- FastAPI backend with a React/TypeScript frontend, hosted on Hugging Face Spaces