Agentic AI Engineer, building AI agents that do real work in production

I'm an Agentic AI engineer. I design and ship LLM systems that run in production under real traffic: agentic RAG with verified citations, real-time voice agents, and resilient multi-provider infrastructure.

Most of my effort goes into the engineering around the model (orchestration, retrieval, evaluation, observability) because that is what turns a good demo into something people rely on. I started in IT security, and it shows in how I build: failure modes and guardrails come first.

The demo is the easy part. An agent that is reliable, evaluated, and safe in production is the real work.

• Agentic RAG in production: multi-mode agent orchestration with parallel retrieval and structured citations, serving real users daily.
• Citation verification (anti-hallucination): every reference checked for existence, accuracy, and validity before display; anything unverifiable is neutralized before it reaches the user.
• Hybrid retrieval at scale: vector + keyword + reranking over 30,000+ domain documents, search latency cut 5x.
• Real-time voice agents: a multi-tenant conversational voice platform (WebRTC, interruption handling), in pilot with a major bank.
• Continuous evaluation: LLM-as-judge on self-hosted Opik, catching quality regressions before every release.
• Performance engineering: critical production endpoints taken from 90 s to 1.9 s response time.

• 🤖 Multi-agent orchestration: planning, tool and function calling, routing, state, and human-in-the-loop.
• 🔎 Agentic search and agentic RAG: query planning, hybrid retrieval, reranking, and validation loops instead of retrieve-once-and-hope.
• 📊 Evaluation and reliability: eval suites, tracing, guardrails, and cost control treated as part of the build, not an afterthought.
• 🌱 Open source: building and contributing in the agents, retrieval, and evals ecosystem.

Plan the query → route to the right source → hybrid retrieval (dense + keyword) → rerank → validate, and retry if needed → answer with verified citations

Agents earn their extra latency and cost only on multi-part, context-dependent questions. When a single call is enough, I keep it simple.

Agents & orchestration  

Retrieval & RAG  

Voice AI  

Evals & observability  

Models  

Backend & data  

• Reliability first: correctness, observability, and failure modes before features.
• Evaluation as infrastructure: checks that gate a release, not a one-off accuracy number.
• Right-sized systems: reach for agents when the problem needs them, not by default.
• Safety by default: guardrails, careful data handling, and responsible release.

I like talking about AI agents, agentic search and RAG, voice agents, and evaluation. Happy to pair, review, or think through a hard problem, and open to collaborating on open-source agentic AI or with teams building agentic products.