Three-plus years working with data taught me one thing: the model is easy; the loop is the product. Pipelines, evals, drift, citations — the stuff that isn’t pretty on a slide is what keeps systems alive in production. Now I’m building agents that plan, query, and cite their own homework.
Ask a question in English. A multi-agent system plans the work, writes SQL, validates the result, renders a chart, and narrates the answer with citations back to the source rows. Built on the NYC Airbnb corpus, designed to plug into any warehouse.
Predictive model over 250M patient-claims identifying BTC patients ~45 days earlier than the standard diagnosis lag. Hybrid clinical rules + K-means/GMM + Transformer NLP clustering on diagnosis narratives. Presented at PMSA 2025.
RAG chatbot answering company financial questions from SEC filings with line-level citations. Semantic retrieval with ChromaDB + text-embedding-3-large; automatic ticker and period parsing; Claude Opus as eval judge.
RAG-augmented classifier categorizing 7M+ farmer transactions across 500+ categories to drive creditworthiness for microloans. Hybrid rule + LLM pipeline reached 99% accuracy on a 10K holdout and cut manual review by 95%.
Professors post a question, students answer via QR, and an LLM summarizes responses into 4–6 themed cards in real time. FastAPI + SSE with a 5-model OpenRouter fallback chain. Single service on Railway.
Chatbot enabling conversations with 60+ historical figures with era-appropriate knowledge boundaries. Dinner-Party mode for 2–5 figure multi-agent dialogue. Flask + SSE streaming, OpenRouter multi-model with fallbacks.
DOM-injected reply generator for X. Three rhetorical angles per request, image-aware context extraction, voice-learning from user selections, streaming responses.
Narrow-scope chatbot that identifies APA 7 / MLA 9 / Chicago 17 violations with rule-IDs and quoted evidence. Vertex AI (Gemini 2.0 Flash Lite) + FastAPI. Three-method eval suite with 30+ test cases.
Fine-grained playback-rate controller that persists per-site preferences and survives segment changes. Works across YouTube, Netflix, Coursera, generic HTML5.