AI Product Architecture & Delivery
A person approves the step, the data has to permit the claim, and there is a record of what actually happened. It is the least impressive part of an AI demo, the only part that decides whether the thing can be used at all, and the part I build first.
Before that I ran delivery for other people — a 40+ person organisation at IOG (Cardano) shipping the Partner Chains SDK, an EVM sidechain and a cross-chain bridge. I still do both halves: architecture in the morning, the deploy playbook in the afternoon.
What each one refuses to do
Process mining for the physical world: it reconstructs how a plant actually runs from the plant's own event log, not from the process diagram someone drew once. The run in the screenshot is a synthetic factory — 3,000 cases, 9,000 events, every one attributable to a case — the benchmark the engine is built against before the first real line.
A correlation is marked candidate and nothing more. It is promoted to confirmed — or refuted — only after someone authorised an intervention and reality answered. Before any of that, an honesty gate decides whether the data even permits the claim. Amber and green are never drawn alike, because one of them is a guess.
A desktop application where agents do the real work of a business process, on a git-versioned workspace — the state of the work is a folder of files you can open and read, not a database you have to trust. Early access is by invitation while it ships to the first design partners.
Deterministic rules keep the agents in bounds, a human approves every step, and git records exactly what happened. No vibes, no black box, no blind trust.
An operator console for dispatch, teleoperation, re-identification and mission control — over CCTV perception and robots on ROS2, with missions dispatched as sequences of primitives over MQTT. Two robots on the fleet today, both simulated. Not my product: I run technical product delivery for the team that builds it. First pilot: a datacenter in Japan, with BlockValue.
The system does not only watch — it sends a robot somewhere. The part gated first is teleoperation: the operator's control lease runs on a deadman timer and drops the robot to zero velocity by itself, two seconds after the operator goes quiet. Every mission the system starts announces itself in the operator feed. Mission approval — the machine asking before it acts — is designed and not yet built; that is the next gate. The failure mode is the robot stopping, not the robot carrying on.
Why I know who to call
At ETH Warsaw 2024 I was one of four speakers on the talk and the hands-on workshop. People walked out with an Ethereum node running on a Raspberry Pi — not with slides about one.
I was product manager on the project. Its entire design goal is lowering a barrier: a full node, set up by someone who is not an infrastructure engineer.
Robotics engineers meet over breakfast, once a month, in Warsaw and Tokyo, with Los Angeles opening this year. The house rule is no sales pitches — researchers and founders show the work.
I build and run the site — the domain, the design and the code are mine. It is the surface the whole thing is found through, and it is where I meet most of the robotics people I end up working with.
Since 2012
Where I have worked — employers and projects, not endorsements
My own practice across applied AI, EVM and DePIN. I set the direction, win the work, build the thing and run the infrastructure — self-hosted, with me on the pager. The systems above were built on an agentic SDLC I run every day: named agent roles with explicit guardrails, specs before code, and a grader I run over the output before I accept it.
A 40+ person cross-functional organisation — engineering, DevOps, QA, release and program management — delivering the Partner Chains SDK, the EVM sidechain and a cross-chain bridge. The code is public: partner-chains and sc-evm. I owned risk, dependencies, the release process, vendor contracts and C-suite reporting.
Built the R&D function from zero — development, DevOps, QA and PMO — then ran it with 30+ people. Two Layer-1 chains and their application ecosystems: Electric Cash and Bitcoin Vault, with wallets, block explorer and mining pools.
My first company, from zero: fundraising, contracts, recruitment, product and delivery.
You get a yes or a no, plus where I think the risk sits.
What I am looking for: one AI system that has to work for real — one at a time. Full-time, fractional at 10–20 hours a week, or a fixed-scope engagement invoiced through my own practice. Remote-first, European hours.
jakubkornafel@gmail.com