About

I help teams make agent systems trustworthy enough to run in production: bounded in what they can touch, honest about when they fail, and measured on the work they actually complete.

Most agent problems I see aren’t model problems. They’re ordinary engineering problems that the model hides: a sandbox that isn’t one, a test that passes without running, a dashboard that reports “no issues” because nobody checked, a cost curve that looks efficient because the job was never finished. Finding those is the work.

Where I’ve done this#

Principal Engineering Researcher, Artificial Quirks (2026–present). I build and stress-test my own agent orchestration and verification platform: isolated execution in Docker and Git worktrees, approval records bound to code changes, scope and secret controls, ledgered token and cost budgets, evaluation scoreboards and OpenTelemetry tracing. Everything on this site comes out of that work, with the numbers and the failures left in.

Lead AI Solution Engineer / Architect, Lloyds Banking Group (2025–2026). I led technical direction for the Conversational Service Team, which provides 24/7 service to millions of customers across Lloyds, Halifax and Bank of Scotland, and its move towards GenAI and agentic solutions. I authored the bank-wide agentic AI platform architecture: use-case value, risks, feasibility, reusable patterns and roadmaps. I also designed stateful multi-agent customer journeys (Python, LangGraph, Google ADK, RAG) with PII scrubbing, input/output validation and deterministic fallbacks. Much of the job was building consensus across engineering, cloud, security and regulatory committees, with decision logs and audit trails to match.

Co-founder, Smart City BG (2024–2025). I took a citizen-facing RAG assistant for municipal administration from user needs to a working prototype, on local, privately hosted LLMs (Ollama, LangChain, LangGraph, ChromaDB).

Quantitative research (2023–2024). I designed and evaluated trading strategies with reinforcement learning and temporal convolutional networks, plus systematic backtesting tools. It taught me to distrust any metric that hasn’t been checked against the outcome it claims to measure, and I apply the same discipline to agent telemetry.

Before AI#

I’ve built software professionally since 2000, from Delphi and .NET desktop applications to cloud-native platforms. The longest thread runs through banking. At Lloyds Banking Group I was Technical Architect and Tech Lead (2016–2019, via Wipro Digital) across cards, mortgages, home insurance and open banking. I came back as Senior Solution Architect (2022–2023, via Sapient) for the Commercial Lending digital onboarding journey. In between and before, I worked as an engineer at the Financial Times, Tesco Technology, Trimble, Alpari and others.

That background matters for agent work. Most of the hard problems are about integrations, permissions, CI/CD, audit and operations, the same things enterprise platforms have always had to get right.

I hold a bachelor’s degree in Mathematics and master’s degrees in Software Technologies and in Economics and Accounting from Plovdiv University.

What I bring to an engagement#

• An outside read. I have no product to sell you and no model vendor to favour.
• Evidence over opinion. Findings come with the run, trace or test behind them, labelled as observation, inference or provisional.
• Architecture and hands-on depth. I’ve written bank-wide AI architecture and I write harness code. I can talk risk with your leadership and read your code with your engineers.
• Regulated-environment experience. I know what security, risk and compliance reviewers need to see, because I’ve had to get agentic designs through them.

Working stack: Python, TypeScript, LangGraph, Google ADK, MCP, RAG and vector stores, GCP (Vertex AI, Cloud Run, Terraform), Azure, AWS, Docker and Kubernetes.

Work with me#

My main engagement is a fixed-scope Agent Runtime & Evaluation Review: two weeks, a written report, ranked fixes. The services page explains how to start. A short written brief is all I need.

I publish new findings on the blog and the Artificial Quirks LinkedIn page ↗.