Engineer · Founder · Vangelis Technologies
Simulation substrates for AI that has to act Most AI systems produce text. The ones that matter next take actions with consequences — in simulation first, then in the physical world. That needs typed state you can query, execution you can replay, and evidence that outlives the run.
| Work | Simulation Sim | Agents Agt | Robotics Rob | Data Data | Infra Inf |
|---|---|---|---|---|---|
| Archetype | |||||
| world-examples | |||||
| awesome-physical-ai | |||||
| sessionscope | |||||
| gepa | |||||
| Claude Code in the Wild: 3 Builder Demos (+ Fable 5) | |||||
| Claude Code and Beyond: State of the Art Agentic Coding Workflows | |||||
| Claude Code and Beyond: 4 Real Agentic Coding Workflows | |||||
| How 2 Engineers Rebuilt a $2M Platform in 2 Months with AI Agents | |||||
| Fever Dreams: Working With and Against AI Sycophancy |
Building the State of the Art is a San Francisco series I started and co-host with Lee of Rally SF. It came out of a plain observation: almost everyone learning these agent tools is doing it alone, behind a laptop. So we put builders in a room and have them show what actually works.
Three builders demo what they actually ship with, plus a first look at Fable 5. Orchestration and harness design carried the night: how much context an agent needs to act usefully, and how to keep the loop verifiable once it does.
The fourth installment, and the longest. Skills dominated — how builders package repeatable procedure so an agent can be trusted with it — alongside review discipline, evals, and the first real comparisons across Claude Code, Codex, and Fable.
Four builders walk through the workflow they genuinely use, end to end. The recurring threads: writing specs an agent can execute, keeping pull requests small enough that a human can still review them, and where token budgets stop being a detail and start dictating architecture.
The series' first full case study: two engineers replace a $2M platform in two months of agent-driven work. Concrete on the parts that usually get waved at — how the work was specified, split into reviewable pull requests, and gated on tests rather than vibes.
Why an assistant's agreeableness is a failure mode rather than a courtesy. The talk walks the good (the barrier to shipping software has collapsed), the bad (human-feedback finetuning that rewards flattery over truth), and the ugly (the GPT-4o sycophancy episode and so-called "ChatGPT psychosis") — then lands on Narcissus as the mental model: a reflection you cannot pull away from, even as it costs you. Position: LLMs are mirrors, not oracles, and every output is arbitrary until vetted.
An AI-native ECS simulation engine: Pydantic's composability for typed components, Daft dataframes for distributed processing, and Iceberg for multi-modal storage. Entities are rows, systems are transformations, and world state is a table you can query after the fact.
Minimal component, prefab, system, and graph composition examples for Archetype — the smallest working thing for each concept, so the engine can be learned by reading one file at a time.
A curated list of open-source projects for Physical AI — robotics, embodied intelligence, simulation, and the infrastructure that makes it work.
See how you actually work. A privacy-first browser analyzer for Claude Code, Codex, and ChatGPT session files — zero upload, runs entirely offline.
Optimizing prompts, code, and more with reflective text evolution — treating the artifact under improvement as something a model can critique and rewrite against a measurable objective.
I work on AI-native simulation — the layer between a model deciding something and that decision having consequences. Archetype, the engine I build at Vangelis Technologies, stores world state as columnar data, so a run can be replayed, queried, and audited rather than merely logged.
Before the state is trustworthy, the execution has to be. Much of the work is unglamorous and load-bearing: typed task graphs, validator-gated transitions, durable checkpoints, evidence that survives failure. I care more about whether a claim is verifiable than whether it is impressive.
Adjacent to that: Physical AI — robotics and embodied intelligence, where simulation stops being an abstraction and becomes the only safe place to be wrong.
Currently
Building Archetype and its runtime at Vangelis Technologies.
Interested in
Simulation substrates, agent execution and evidence, robot policy evaluation, columnar data systems, and knowledge systems that compound rather than decay.
Elsewhere