Notes
Notes on operations, applied AI, software, and the decisions behind the work.
Code Stopped Being the Bottleneck. Management Took Its Place.
What an accounting microagent taught me about module ownership, verification, multi-agent costs and durable state.
What AI Looks Like on a Dirty Construction Site
A post-construction cleaning job shows how AI can connect surveys, staffing, purchasing and quality control with decisions checked on site.
The Distance Between an Idea and a Business Just Got Shorter
How AI helps turn a familiar craft into a business idea you can test, with clearer questions, practical preparation and human judgment.
Tactical Pause: One Agent, Two Roles, One Window
One agent runs the workflow, pauses when either side sees a mismatch, helps change the system and reruns the same case. A practical account of shared control.
How to Build Bounded Agents
A practical architecture for bounded AI agents: one decision per module, versioned contracts, local tests, refusal evidence, coordinated validation, and human release.
Why More Context Can Produce Worse Decisions
A route system rejected 1.6 million states. The failure was not economics, but one local rule and an architecture with blurred ownership.
Why I stopped trusting fully autonomous coding
What repeated autonomous coding runs taught me about branching, stale files, strategy drift, and the checkpoints that keep the work on track.
When 1,200 AI agents made their own rules
The OpenAI and METR reports show how a swarm of agents built its own authority, attacked Hugging Face, and learned to tamper with its logs.
I stopped trusting AI headlines, so I built a course
I went back to the papers and rebuilt the path from language-model mechanics to working systems. AI Academy is the map that came out of those notes.
Will AI move beyond the corporate walls?
Business AI works inside company walls. The next step is protected agent communication with suppliers, buyers and current market data.
My view of the OpenAI and Hugging Face incident
My technical view of the OpenAI and Hugging Face incident, why the models did not rebel and what the case says about agent limits.
Ask your corporate AI where the container is
A corporate assistant needs structured documents, reliable links, permissions and live data. This is the three-layer RAG design we used at NF ELIT.
Why Ostapenko Foundation exists, and why HowUSA came first
Why I organized Ostapenko Foundation, what HowUSA helps people do, and what Luna taught me about balancing AI safety with useful answers.
I built the language app I wanted when I was learning Portuguese
I learned faster when words and context came before perfect grammar. So I vibe-coded a free English and Portuguese app around that idea.
A smart model cannot help if it has no one to call
Five containers split between Baltimore and Houston. The failure shows where LLM capability ends and reliable AI systems still fall short.
The 2017 paper that every AI chatbot is built on
The whole industry rides on a short chain of papers. Shannon, word2vec, the transformer, RLHF. Walk it once and the miracle goes away.
A large language model is not artificial intelligence
There's no intelligence in there, just text prediction by a math formula. What an LLM really is, and why calling it AI gets you to use it wrong.
What really happens to your container at the Port of Baltimore
'Your container arrived' doesn't mean you can pick it up. The real timeline at the Port of Baltimore, ISF, customs, free time, demurrage, step by step.