Projects
Five systems, written up properly.
Mechanism first, then the number it produced. Each write-up says what the system had to survive, how it is built, and how I know it works — the same order I would defend it in an interview.
Corrective RAG with self-reflective retrieval
Most retrieval-augmented systems fail quietly. This one is built so that when retrieval goes wrong, the system notices, tries again a different way, and refuses rather than inventing an answer.
LangGraph
GPT-4o
QdrantCross-EncoderReal-time voice AI agent integration
In text, a slow model is a spinner. On a phone call it is silence, and silence is the one thing a caller reads as failure. Everything about this system is shaped by a latency budget you cannot negotiate with.
Telnyx
GPT-4RetellVapiMulti-agent content platform (planner → critic loop)
Calling something multi-agent usually means a prompt was split into four prompts. What makes this a graph is the edge that goes backwards: the critic can reject work and send it back to be done again.
LangGraph
GPT-4o
Claude
FastAPI
Next.js
Make.comAI-powered HRMS with semantic policy search
Every company has the answer written down somewhere. The problem is that it is on page 34 of a PDF nobody opens, and so the question gets asked to a person instead.
LangChain
HuggingFace
React
PostgreSQLRetrieval & grounding evaluation harness
Anyone can demo a RAG pipeline. The hard part is proving it got better — and this is where the 72% to 89% number came from, a measurement loop rather than a guess.
PythonGolden setsLLM-as-judgeGroundedness scoring
GitHub Actions