Validated RAG Eval Harness
A human-validated comparison of three LLM-as-judge prompts for answer correctness and citation support, with Cohen's kappa, question-type disagreements, and labeling costs.
Projects
Tools and experiments exploring how to engineer AI systems that are measurable, useful, and responsible in practice.
Chunking Strategy Ablation
A controlled comparison of fixed, recursive, sentence-window, and semantic chunking on retrieval, context tokens, answer accuracy, and latency.
Sequence Models at Small Scale
A parameter-matched comparison of RNN, LSTM, GRU, and a tiny Transformer on character prediction and 100-step recall.
Tokenizer Ablation
A byte-level BPE tokenizer trained from scratch and evaluated against downstream code-identifier accuracy under fixed token and byte budgets.
Learned ANN Index
An experimental recursive partition model for vector search, compared with IVF-PQ, HNSW, and ScaNN on recall, throughput, build time, and index memory.
Chat with My Research
A citation-first, LangGraph RAG application that decomposes multi-hop questions with MCTS, compares search strategies, retrieves local source passages, and evaluates citations.
Personal Chief of Staff
A local-first, LangGraph-supervised assistant for daily briefings, email triage, meeting prep, and scheduling suggestions through read-only MCP servers.
BenchmarkKit
An open-source toolkit for measuring LLM inference and RAG latency, throughput, cost, retrieval recall, and answer support.