Work
Three pieces I'd want read first, then the earlier systems that taught me the parts I now take for granted. No mock-ups — each entry is something that was built and used.
Featured
Account questions across chat and voice. I built the routing, approved-tool layer and Azure deployment, with LangGraph for questions that need multiple lookups.
A monthly budget becomes a daily allowance. I designed, built and operate the app, including a text-and-voice assistant that asks before changing data.
A portfolio should answer follow-up questions. I built this assistant to answer from my published work, cite sources and say when the evidence is missing.
The corpus is small enough that a top-K cutoff turned out to be a liability — the first version dropped the one chunk a recruiter's question needed. It now gives the model the whole corpus and uses the vector scores to rank citations against the answer that was actually written, which is a more honest source trail than similarity to the question.
Earlier work
Document automation
Invoices, tax documents, scanned PDFs and spam arriving by email, routed into ERP workflows: attachment rendering, OCR extraction with fallbacks, AI classification with confidence scoring, Azure SQL logging, and a manual-review path for anything below threshold.
Research intelligence
A workspace for exploring scientific papers: PDF parsing, Ollama/Phi-3 retrieval chat, embedding-based paper comparison, PubMed retrieval and relationship analysis with NetworkX.
Search
Spell correction and query cleanup, semantic matching, feature extraction and product recommendation over a catalogue — a first pass at the "understand what they meant, not what they typed" problem the telecom router later solved properly.
Security
Secret keys generated from chaotic sequences for unpredictability, dynamic fingerprint selection (one of ten stored samples chosen by generated index), and layered authentication for sensitive transactions.