I build AI features that have to work as part of a larger software system. My recent work has focused on document intelligence, structured extraction, retrieval, and agent workflows. I came to this through backend and data engineering, so I tend to think about the APIs, state, and checks around a model as carefully as the model call itself.
My work has moved from Python, ETL, and APIs toward LLM applications and, more recently, stateful agent systems. That progression still shapes how I build: understand the workflow first, use a model where it helps, and make the result clear enough to validate and use downstream.
Built stages for extracting, validating, and enriching information from complex insurance documents. The workflow combined LangGraph, structured outputs, and deterministic checks to catch missing, malformed, or conflicting values before they moved downstream.
Worked on a proof of concept for extracting product attributes through coordinated agent steps. The design included structured results, queued processing, and a human review path for uncertain cases.
Developed an AI workflow around automotive parts information, bringing together retrieval, semantic search, and image grading to help organize and assess product data.
Earlier projects included document intelligence and RAG, invoice extraction, and bank-statement processing. These systems paired Python APIs and data pipelines with language models to turn unstructured documents into information that business workflows could use.
Built LLM-based document question-answering and conversational applications, along with an early voice-to-email workflow that turned spoken instructions into draft messages.
Project summaries are intentionally general and omit client-confidential details.
- Start with the task, the data it depends on, and what can go wrong.
- Give model steps clear inputs and structured outputs.
- Use ordinary code for rules that should be predictable.
- Make uncertain results reviewable by a person.
- Treat the API and the downstream handoff as part of the AI feature.
Python · FastAPI · LangGraph · LangChain · RAG · OpenAI · Azure OpenAI · Pydantic · MongoDB · PostgreSQL · Azure · AWS
- Document extraction workbench — extract information from PDFs and images, review it, and export structured CSV or Excel data.
- AI-assisted ETL with a chatbot — extract information from unstructured documents, store it, and query it conversationally.
- Voice-based email helper — turn spoken, multilingual instructions into email drafts.
The live portfolio brings together selected work, side projects, my engineering approach, and career history.