Challenge

While working on a POC with an AI hardware vendor to evaluate local inference capabilities, the limitations of basic RAG prototypes became clear — they weren't portable, weren't production-ready, and couldn't be reliably deployed anywhere outside a controlled environment.

Solution

Took the POC as a springboard to build a genuinely production-grade, deployable-anywhere RAG system. Used FastAPI and LangChain with deep agent orchestration, Qdrant for vector search, and PostgreSQL for persistent state. Deployed on AWS EC2 to validate real-world deployability, with LangSmith for full observability.

Challenge

Pure LLM coding suggestions often lack local file context and execution capability, leading to generic and often broken code in complex projects.

Solution

Built a VS Code extension with a LangGraph ReAct agent that performs autonomous file operations (read, write, edit) with built-in AI reasoning and a visual diff interface.

Challenge

Customer support teams were overwhelmed by repetitive queries buried in static documentation.

Solution

Designed end-to-end RAG pipeline with document processing and semantic search. Integrated Flowise workflows with Supabase and Pinecone vector databases. Built web interface using Lovable (v0-style AI builder). Implemented LLM observability with LangSmith. Achieved 2nd place out of 6 teams.

Challenge

EJ files were stored locally on ATMs, requiring manual transfer and causing 24-hour reconciliation delays and high dispute failure rates.

Solution

Designed a 4-component pipeline (EJLIVE, EJSYNC Client/Server/Web) using Python/Django and React to automate log collection and provide a central search interface.

Challenge

ATM card request files were EMV-encoded and unreadable, causing slow duplicate detection and frequent production errors across 700 branches.

Solution

Developed a cross-platform desktop application using Python and PySide6 that parses EMV standards, maps branches, and flags duplicate requests automatically.