Retrieval, agents and evaluation: small working systems you can poke at, running live in your browser.
Ask a question. BM25 answers instantly; load the embedding model and results upgrade to hybrid search, fused with reciprocal rank fusion. The map shows where every passage, and your query, lands in vector space.
An agent is only as debuggable as its traces. Replay three runs (a RAG copilot, a data-incident agent and a multi-agent research swarm) and inspect every LLM call, tool call, retry and eval.
Demo traces: synthetic runs modelled on the span structure of real production traces.
AI engineering work, newest first.
BM25 and MiniLM embeddings fused with RRF, served from a static site. A Web Worker keeps the interface smooth while the model indexes passages.
A zero-dependency replay of agent traces with a span waterfall, a payload inspector and streamed outputs. It is the debugging loop every agent deserves.
Retrieval-augmented chatbot on a Weaviate vector store. It indexes any document corpus and answers with citations, with pluggable LLM backends.
LLM pipeline on Vertex AI that researches keywords, drafts copy and scores readability against competing pages.
A machine-readable map of this website for language models and AI crawlers: the robots.txt of the agent era.