Senior Principal Software Engineer with 19+ years in Java — most recently 13 years at Oracle, building enterprise platform infrastructure. Based in Prague.
I work end-to-end across the stack but my strongest zone is the middle of it: real-time search and indexing, pluggable persistence, custom compile-time frameworks, event-driven integration over Jakarta EE.
Coherence over orthodoxy: a system is valid if it is internally consistent, falsifiable, and explanatory, regardless of its alignment with current dogma.
🔧 tiko-di — Compile-time dependency injection framework for Java 21+. All wiring validated and generated by an annotation processor; no reflection, no classpath scanning, no runtime DI surprises. Ships with a built-in event bus that swaps transparently between local and distributed implementations. Deliberately built to be AI-buildable without prior model familiarity — it brings its own machine-readable in-repo guidance, compile-time wiring validation, and an MCP topology server, rather than relying on a model having seen it in training (see the benchmark below).
📊 llm-framework-benchmark — External-oracle-graded benchmark of how well AI coding agents build the same framework-neutral spec on different stacks, run across three models (Claude Sonnet 4.6 / Fable 5 / Opus 4.8, n=5). It surfaced a sharp, model-dependent version-recency effect — a new Spring Boot major (4.0.6) was a near-total wall for Sonnet 4.6 (median 0%) yet solved by Fable 5 and Opus 4.8 (median 100%) — while tiko, absent from training data, reached 86–100% median through its in-repo affordances.
🧠 trie-memory — Experimental MCP memory server in Rust. Append-only hierarchical memory with explicit epistemic states (Answer / Partial / Unknown) — designed not to fabricate when it doesn't actually remember. Research PoC built on a delta-encoded recognition trie paired with a concept store.
This is my professional portfolio. For experimental work, prototypes, and side investigations across physics, linguistics, and AI tooling, see @jerry-samek.