I design and ship agentic AI systems — combining LLMs with agentic retrieval, knowledge graphs and continuous evaluation — that deliver trustworthy, evidence-based results in high-stakes, regulated domains — healthcare, finance, legal and enterprise knowledge.
SCROLL TO EXPLORELarge Language Models are trained to predict text, not truth. They can be confidently wrong, especially when the cost of error is high.
Even 2026 frontier models hallucinate 4–19% depending on the task — citation accuracy is the worst offender.
Built-in search is generic. Your domain knowledge isn’t in the weights — it needs governed, traceable retrieval.
The model can’t check its answer against your data, policies and sources. That check has to be engineered.
Prompt injection, over-permissive tools, irreversible actions. Agents need sandboxing, permissioning and human-in-the-loop.
WHY IT HAPPENS
Models learn P(text), but we need P(true | text). These are not the same.
Even a perfect model will hallucinate because of uncertainty, compression and noisy data.
Model optimizes
P(text)
We need
P(true | text)
P(hallucination) > 0
Always hallucinate.
The world is complex. Context is limited.
Models compress knowledge.
Compression creates uncertainty.
Uncertainty leads to hallucinations.
THE SOLUTION
I build agentic architectures where the LLM is just one component. The agent plans, retrieves, self-checks and retries — every answer is grounded in data, verified, and fully traceable.
The agent decomposes the task and searches trusted sources and tools (MCP)
Knowledge graph links entities, relations and rules — curated context, not a dump
Self-check loop: fact-checking, contradiction detection, corrective re-retrieval
Least-privilege tools, sandboxed execution, prompt-injection defense, human approval for irreversible actions
LLM generates answers grounded in validated evidence — with guardrails
Citations, confidence scores, and full tracing of every step and tool call
BUSINESS IMPACT
Across deployed client systems — case studies available on request.
−60–70%
hallucination rate
(in real systems)
+30–50%
answer accuracy
and relevance
100%
traceable answers
with sources
∞
built to scale
with your data
PRODUCTION READINESS
An agent in production is a security surface and an operational system — not a demo. I design for both from day one.
Aligned with the OWASP Top 10 for Agentic Applications (2026).
Evals practice follows Anthropic’s guidance on agent evals (2026).
ABOUT ME
Staff AI Engineer / AI Architect
I design and develop production-ready AI systems for regulated, high-stakes domains where accuracy and trust matter most.
Single- and multi-agent systems, orchestration, memory design
Python, FastAPI, LangGraph, MCP integrations, Docker
Agentic RAG, knowledge graphs, context engineering
Evals, tracing & observability, guardrails
