AI engineering, from CUDA kernels to agentic pipelines.

Built a RAG reranking pipeline (SciFact/BEIR, MiniLM cross-encoder), profiled it with Nsight Compute to find the true attention bottleneck, then hand-wrote a fused CUDA kernel — register caching, shared-memory tiling, float4 vectorization, and an online-softmax redesign — cutting latency from 286ms to 10.5ms, roughly 29x faster than the naive baseline.

Added padding-mask support and a tile-skip optimization guided by Nsight stall analysis, reaching 7.70ms against PyTorch SDPA's 10.27ms masked latency — about 25% faster. Integrated the kernel into an unmodified HuggingFace BERT model via transformers' AttentionInterface, validating near machine-precision logits (1.9e-6 error) and zero ranking regression (20/20 top-10 match).

Answers questions about an unfamiliar codebase the way a senior engineer onboarding you would: by grading its own retrieved context and correcting course when it's wrong. Architected as an 8-node Corrective RAG (CRAG) LangGraph StateGraph — documents are graded correct, ambiguous, or incorrect, and anything short of "correct" routes to query refinement, a live web fallback, or reranking before generation, so a weak retrieval doesn't quietly turn into a hallucinated answer.

Built a custom asymmetric MMR retriever and a structured-output relevance filter after root-causing a production LLMChainFilter failure caused by Gemini's multi-block response format. Every citation is checked against the real evidence set by a custom guardrail with two independent retry-loop counters, one for retrieval and one for generation. Full pipeline evaluated with RAGAS: 0.884 Faithfulness, 0.839 Context Precision, sliced per graph execution path.

Built for an AI Engineer technical assignment: upload any CSV, ask questions in plain language, and get back real SQL queries executed against an actual database, plus charts, anomaly detection, data-quality checks, and forecasting. Structured as an explicit multi-node LangGraph with a router, rather than one prompt trying to do everything.

Built for the ByteVox AI/ML Engineering Internship assignment. Combines hybrid retrieval (BM25 + embeddings) with LLM-based reranking, plus an agentic retrieval layer built with LangGraph as the bonus track.

Built for Think9's AI & Intelligence Challenge: an agentic system that produces a structured launch blueprint — market read, positioning, pricing, sourcing shortlist, creative brief — for new consumer brands, drawing on a memory layer built from past launches. Built as a full 4-agent LangGraph pipeline (market signal, memory/RAG, synthesis, and critic agents) rather than a cut-down demo, on a 3-day deadline.