Anton Dziatkovskii

AI-agent memory, verification and operations, with public evidence.

I have a brain tumor. I found it ten years ago, and that is why I started building a second brain long before LLMs became a hype: my own notes, indexed and searchable, retrieval over a personal archive. Today that archive feeds my agents. Agent memory is my personal pain, not a product idea. PhD, Python/C++. Seeking Research Engineer roles.

Technical lead at Palo Alto AI Research Lab. The work covers persistent memory, deterministic checks and coordination across machines. Selected projects and their evidence · Upstream contributions with actual status.

CS security (MEPhI), Python and C++, and hired operating leadership (COO/CTO) at a venture incubator since 2015. Current preprints, experiments and experience reports document the methods and their limits.

Off the clock: triathlon (mid-pack, drifting to the back, honest about it), surfing, and a licenced helicopter pilot now training on fixed-wing — because I build bunkers. Several are already standing in New Zealand and Australia, and flying is how you reach a shelter when things go sideways.

Reproduce the system in 15 min → claw-consensus (MIT · offline demo · 11 documented failure modes) Read the method → preprints GitHub → all repos Academic profile → publications & research

📄 Resume — PDF (ATS-friendly) · Resume — JSON · 🎓 Academic profile · All publications · 🔧 Join the fleet · 🔢 Fleet token meter · Open-source contributions — 251 pull requests across 165 repos, each with its real status

Proof of work (flagships, not a gallery)

claw-consensus — multi-machine agent consensus you can reproduce in 15 minutes.
Offline demo, published evals, honest FAILURE-MODES.md (11 documented failure regimes). Theory shipped to prod.
Real output of python demo/demo.py — captured 2026-07-22, stock Windows laptop, no network, no API key.
verbatim-citation-gate — catch fabricated RAG citations before they reach the user.
A zero-token verbatim gate plus a burden-of-proof judge, framework-agnostic. Open source, MIT. Docs — including the section that names the case where it still fails.
claude-bible — the governance layer: one rulebook that an operator, his agents and his team all load and obey.
Rules as files with precedence and a routing tree, a declined-decisions journal, objection sparring. Most-starred repo of the lab. Docs.
Upstream, in code I do not own — security and reliability fixes merged by the maintainers of projects I do not control.
Most recent: microsoft/semantic-kernel #14371 (merged 5 Oct 2026) closed a check-time/use-time gap in the Python OpenAPI plugin's anti-SSRF validator — the request now goes to the exact address the validator vetted, so on the plugin's built-in HTTP client a DNS-rebinding hostname can no longer pass the check and then reach an internal address. Earlier: UK AISI inspect_ai #4769 (a grader panel decided by strict majority, not mode) and google-gemini/cookbook #1296. Every pull request with its real status →
The method behind it — ten preprints, each published with a DOI (Zenodo, CC BY 4.0). The four core ones: Operating a Human-Governed Multi-Machine LLM Agent Fleet (the experience report behind claw-consensus), Homeostatic Governance (three-circuit control: main / adaptive / protective), When Does Graph Expansion Help Personal-Corpus Retrieval? (v0.3: the protocol run on 38 queries; graph expansion did not help, an honest negative result behind sqlite-graph-memory), and Le Chatelier Tokenomics (multi-asset coupled equilibria). Five more on running agents in production: Keeping Agents on a Leash (an eight-domain control model), Rules as Files (governing agents across sessions and machines), Silent by Construction (a casebook of control-plane failures), Instructions That Never Arrive (do compaction instructions reach the summarizer?) and Facts Before Explanation (deterministic rules decide, the model only explains).
The repos are the reference implementations of the papers. The Zenodo record is the version of record; copies anywhere else are mirrors of it. All ten 2026 preprints → · Full list →
相棒 AIBŌ · The Partner — a book written from inside a human+AI company, documenting the method day by day.

How this was built: with Claude as implementation collaborator — and that is the point. I retain full ownership of problem framing, architecture, evaluation and final QA; the repos publish the evals, the failure modes and the daily log, so you can audit the division of labor yourself. Metrics are published or absent.

Research (selected, peer-reviewed)

All 50+ publications → · 138 citations, h-index 7 (Google Scholar, verified 26 Oct 2026)

Developer education track record

Writing & speaking

HackerNoon column (34 essays: DeFi mechanics, CBDC, NFT gas optimization) · interviews in Japanese crypto press (Crypto Times ×3) · 50+ academic publications, PhD in Education (IT) — full academic profile at /scholar/.

Reach

Based in Palo Alto, CA and Sintra, Portugal (WET/WEST), 15 years lived across APAC (Japan, Korea, Thailand, Malaysia, Singapore, HK, Australia). Working languages: EN · RU · Ukrainian · Polish — a live bridge to the Eastern-European engineering diaspora. Work authorization: US O-1 visa (active) and EU citizenship (Polish passport) — no visa sponsorship required in the US or the EU.

Contact

[email protected] Calendly WhatsApp

Find me on

GitHub X Google Scholar ORCID Academia.edu Wellfound

Anton Dziatkovskii · Candidate one-pager · Speaker · Resume (PDF, ATS) · Resume (JSON) · Academic profile · All publications · Preprints 2026 · Open-source contributions · Fleet token meter · Writing · Rus · GitHub
READ THIS WITH AI

One click and an agent reads the article, pulls out the patterns and helps you apply them to your own work.

Codex → ChatGPT → Claude →

— TonyDzi, Palo Alto AI Research Lab · second brain, agent coordination, persistent memory: github.com/tonydzi