Projects
AnswerTrail
deep-research agent RAG retrieval LLM evaluation AWS
A deep-research agent that answers questions over one AI/LLM YouTube channel’s transcripts, with every citation validated against the corpus before an answer is returned. A hand-written ReAct loop searches dense retrieval over Postgres, the whole system deploys live on AWS from a single-command Terraform stack, and a nugget-based evaluation with a cross-family judge calibrated to Cohen’s kappa 0.879 measures how well it answers.
Clinical documentation pipeline
LLM workflow structured extraction clinical NLP
A multi-step LLM workflow that turns a clinician-patient encounter transcript into a structured clinical note. Parallel extraction calls produce typed, metadata-tagged frames that a downstream routing layer places into the correct note sections. Clinician evaluation showed 2× recall of medical facts at maintained accuracy, improved clinical utility over the prior production system, and ~50% better run-to-run consistency on extracted medical problems and clinical facts.
NLPLingo
event extraction causal relations deep learning DARPA IARPA
Raytheon BBN’s deep-learning toolkit for event extraction and causal relation extraction. I created, led, and was the principal developer and maintainer of NLPLingo over its lifecycle, mentoring the scientists and engineers across Raytheon BBN who built on top of it.
The toolkit supported ~$25M in DARPA and IARPA contract value, most notably DARPA World Modelers, where I served as Principal Investigator on the Raytheon BBN team (DTIC final report).
TNLP
encoders decoders information extraction
A single codebase I designed and wrote by hand, covering the model and training stack end to end: sequence and token classification, span-pair relation extraction, seq2seq, contrastive retrieval, instruction and chat fine-tuning, DPO, and end-to-end RAG with a jointly trained retriever and generator. It uses standard backbones (BERT, DeBERTa, T5, LLaMA, Mistral) and the usual libraries where they fit, with custom model and training code where the task needed it: the span-pair model, the triplet contrastive model, and the joint retriever-generator RAG loss are written from scratch, not pulled from a library.
Paper notes archive
75 explainer posts working through the transformer literature one paper at a time: encoders and decoders, efficiency methods, retrieval and RAG, evaluation. Written over about a year, it predates this site and is encyclopedic where this one is argued, kept as a standing reference rather than a current feed.