let current_focus = vec![
    "Evals-first agent development (golden sets, regression gates)",
    "AI agent platforms & orchestration",
    "MCP servers & Claude Code toolkits (skills, subagents, plugins)",
    "Knowledge graphs for agent memory & retrieval",
    "LLM routing, cost & latency engineering (AI FinOps)",
    "High-performance trading systems",
];

Current:

• 🏦 AI Engineering Team Lead @ EuroWAG (fintech, fleet payments): touchless invoice pairing, an AI developer platform used by ~100 engineers, AI-driven FinOps
• 🤖 Founder @ TradeGuard: AI trading platform + Moira, an eval harness and Agent Factory for trading agents
• 🎓 Postdoctoral Researcher (part-time) @ VŠE Prague: Human-AI Collaboration; PhD defended there Oct 2025

Previously:

• 📡 Amdocs: built aOS, an agentic platform framework; replaced an LLM retrieval step with deterministic graph queries (precision/recall ~50% → ~95%, median latency 30 s → 1.1 s)

Agent Orchestration:

• MCP (Model Context Protocol) servers
• Claude Code skills, subagents & plugins
• LangChain & LangGraph workflows
• Mixture-of-Agents / agent boards, Agent Factory pattern
• Claude Code, Cursor, Codex, Gemini CLI

Evals & Quality:

• Golden sets per agent, cross-vendor LLM judging
• Retrieval precision/recall, hallucination rate, provenance
• Cost & latency budgets (median / p95)

LLM & Embeddings:

• Anthropic Claude (primary) • Kimi • OpenAI APIs
• Azure AI Foundry • AWS Bedrock
• HuggingFace Transformers • local inference (Ollama, llama.cpp)

Retrieval & Knowledge:

• Neo4j / Cypher knowledge graphs
• pgvectorscale + DiskANN (PostgreSQL)
• Hybrid search (BM25 + vector)

Green-field agentic platform (agent board, planner, clerk, knowledge-graph memory). The user-facing agents scored well on golden sets but underperformed in the application, so I wrote golden sets for the retrieval layer itself. The evidence said the LLM "data librarian" was the bottleneck; I replaced it with deterministic, parameterised Cypher behind a deliberately narrow MCP tool, plus new KG schemas and indexes.

// Illustrative: the MCP tool accepts only a search term and optional hops
CALL db.index.fulltext.queryNodes('entityIndex', $term) YIELD node AS e, score
CALL apoc.path.subgraphNodes(e, {maxLevel: $hops}) YIELD node AS ctx
RETURN e, score, collect(DISTINCT ctx) AS context
ORDER BY score DESC LIMIT $k

_{Latency includes a Haiku call interpreting results. Cost fell because retry loops between clerk, planner and agent board disappeared.}

Then built the Agent Factory: consultants create new agents over synthetic data without touching code, and each agent ships with its own auto-generated golden set.

# Each new agent gets its own golden set, generated and judged
# by a *different* LLM vendor than the one that built it.
golden = generator.make_questions(agent.spec, n=20, difficulty="graded")
while True:
    report = judge.score(agent, golden, criteria=[
        "retrieval_accuracy", "hallucination_rate",
        "provenance", "cost", "latency",
    ])
    if report.passes(thresholds):
        break
    agent.prompt = judge.propose_revision(agent.prompt, report)

An eval harness for training and evaluating "market wizard" trading agents, executed through the TradeGuard platform:

• Schools: each LLM vendor raises its own agents; another vendor and a human evaluate them
• Lifecycle as phases: ingest → design (Agent Factory) → backtest → forward-test → paper trading → live → archive
• World model in knowledge graphs; agents coordinate through a session KG
• Auditability, tracing and provenance as first-class requirements

# Fuzzy scoring + LLM top-up; a human is called only for
# anomalies the pipeline cannot resolve on its own.
class InvoicePairing:
    def pair(self, vendor_tx, invoice_items):
        match = self.matching_engine.score(vendor_tx, invoice_items)
        if match.resolved:
            return AutoPair(match)
        resolved = self.llm_resolver.try_resolve(match)
        return resolved or HumanReview(match.anomaly)

• 300 fuel vendors onboarded to fully touchless pairing
• Vendor verification with separated reference / hidden sets, regression runs and production confirmation across billing cycles
• KPIs: touchless share, matching coverage vs pair correctness, LLM top-up coverage, escalation rate

• Shared .claude/ toolkit: commands, skills, subagents, MCP servers and setup guides, used by ~100 people across two engineering units; team leads report 10–20% faster delivery
• AutoDoc: AI-generated, human-verified documentation for a 19-service domain, with HITL editing
• Enterprise knowledge base aggregated for agent context
• Founded AI Guild and hands-on AI Lab (biweekly)

• Production Log Analytics spend €12.4k → €1.7k / month (−86.5%), confirmed on billed invoices; ~€130k / year saved
• Follow-up campaign on shared platform subscriptions

Languages:

Rust      ████████████░░  Expert (systems, trading)
Java      ████████████████ Expert (16+ years)
Kotlin    ████████████░░  Advanced
Python    ████████████░░  Advanced (ML/AI)
TypeScript████████░░░░░░  Proficient
Cypher    ██████████░░░░  Advanced (knowledge graphs)

Backend & Microservices:

• Spring Boot • Quarkus 3.x • Micronaut
• Apache Kafka • CDC (Debezium)
• gRPC • GraphQL • REST • WebSockets
• Event-driven (CQRS, Saga, Outbox)

Cloud & DevOps:

• Kubernetes • Docker • Helm
• Azure (expert) • AWS • GCP
• Pulumi IaC • GitLab CI/CD
• Prometheus • Grafana • ELK • Honeycomb

Data:

• PostgreSQL (expert) • Neo4j • Cosmos DB • MongoDB
• Redis • Elasticsearch
• jOOQ • Hibernate • R2DBC

Agentic AI:

Production Systems:

Research & Open Source:

PhD Dissertation (Defended October 2025, VŠE Prague): "Human-AI Programming Role Optimization: Developing a Self-Determination Framework"

Key Finding: AI-assisted development increases programmer motivation by 23–65% when optimized for individual personality types (Big Five) and working styles (Self-Determination Theory).

Practical Applications:

• 🎯 When AI agents should lead vs. support developer decisions
• 🖥️ Designing interfaces that respect developer autonomy
• 📊 Measuring AI tool effectiveness beyond productivity metrics
• 🏢 Change management for AI adoption in enterprise (applied daily via AI Guild / AI Lab)

Select Publications:

• PeerJ CS (Q1): Personality-Driven Pair Programming
• IEEE ICSME (CORE-A): AI-Assisted Programming Psychology (45 citations)
• EASE (CORE-A): Psychological Aspects of Pair Programming
• ACIE'25: Blockchain-Driven Transparent Research (Best Paper)
• CIMPS'22: (Best Paper)
• DD FIS VSE'22: (Best Paper)

Oracle / Eclipse Foundation:

• Designed and implemented Bean Validation (JSR 303) integration on both sides of Java XML binding:
  • JAXB: BV support in the XJC and JXC plugins; BV-annotation package shipped in the JDK 9 distribution
  • EclipseLink MOXy: released in EclipseLink 2.6 (2015)
• 57–92% performance improvements
• 50+ merged PRs

Interests:

• AI agent tooling & evals
• MCP ecosystem
• Rust systems programming

16+ Years Building:

• 200+ microservices in production
• Systems serving millions of users
• Fintech, telecom, CRM, IoT, trading platforms

Where I've Built:

• 🏦 EuroWAG (fleet payments)
• 📡 Amdocs (agentic platforms)
• 📱 T-Mobile Czech Republic
• ☕ Oracle Corporation
• 🏦 Home Credit International
• 🏭 Rockwell Automation
• 🏛️ Ministry of Interior CZ
• 💳 DNZ Finance (crypto)
• 📊 Adastra

Building:

• Evals-first agent harnesses
• MCP servers for enterprise AI
• Knowledge-graph agent memory
• Rust + Python hybrid systems

Exploring:

• Measuring AI benefit via unit economics (cost per invoice / release / incident)
• AI agent payment protocols
• Autonomous agent orchestration
• Browser automation + AI