Projects
Each project below is a full system: ingestion to retrieval to evaluation, or agent to tools to traces. The engineering scope matters more than the screenshot, so that is what I document.
Teams ask questions; their answers are buried in hundreds of PDFs.
Contextual RAG pipeline with PDF ingestion, chunking, Chroma vector search, and hybrid retrieval. Answers ship with citations, and Ragas evals feed a quality dashboard so retrieval regressions are caught, not guessed.
One agent can answer a question. Real workflows need several that cooperate.
Supervisor and specialist agents collaborating through shared state, checkpoints, and human-in-the-loop approval. Every run produces a full multi-agent trace, so behavior is debuggable instead of magical.
Support teams drown in repetitive tickets that follow known playbooks.
Tool-calling agent that searches docs, summarizes issues, and creates tickets, all behind guardrails, tracing, and a human approval step for anything irreversible.
Answers live half in documents, half in database rows. Most copilots pick one.
Hybrid answers across documents and SQL with record-level citations, row-level access control, and governance reports. Data-agent workflows that respect who is asking.
Agents with unrestricted tools are a security incident waiting to happen.
A secure tool registry with permission levels, a local tool server, database query tools, audit logs, and security tests. MCP concepts applied to enterprise constraints.
Agents forget everything, or worse, remember things users never agreed to.
Thread memory, durable facts, and semantic retrieval with user-controlled list/delete flows and privacy audit logic. Memory as a feature users can inspect, not a black box.
LLM apps fail quietly: costs drift, evals decay, tools break overnight.
LLMOps dashboard tracking cost, latency, eval scores, and tool failures, with provider routing, nightly evals, and incident notes. Operational visibility for AI systems.
Before AI
The AI work sits on top of a long full-stack track record: 125+ client projects on Fiverr (Level 2 seller, 132 five-star reviews), 16 Upwork engagements, a production LMS built at Rigrex with Next.js, Prisma, and PostgreSQL, and MVP work for European startups at WeMakeApp. E-commerce, Web3 frontends, dashboards, SaaS billing: the unglamorous plumbing that makes products real.