All posts

A technical account of multilingual RAG evaluation, unsupported questions, and why a refusal phrase cannot prove the assistant stopped.

Three models found the expected documents. Unsupported questions exposed a separate requirement: stop when the knowledge base has no answer.

Our OpenHands-based cloud environment centralizes agent configuration and repeats review and fixes before human approval.

I built a food tracker with AI to fit my own needs. I think SaaS products will need to solve harder problems as more people can build their own alternatives.

Upload a sample PDF, get a drafted accessibility template. Available today in DocEngine, with two non-negotiables built in.

We tried drafting accessibility templates without an LLM first. What shipped gives the model meaning but never a coordinate.

I burned 39% of a weekly Codex quota in four and a half hours. A context-window override I set during GPT-5.5 was still in my config.

After a 14-hour Codex run, I want long agent tasks controlled by a queue with a clear finish line and reviewable evidence.

AI agents increase code output faster than teams increase review capacity. Senior attention moves to understanding, verification, and ownership.

Claude Fable 5 refused a defensive review of our own SaaS application twice, including a prompt that explicitly banned exploit guidance.

DocEngine now runs veraPDF after processing a document and reports PDF/UA-1 conformance and failed checks in the same workflow.

My son wanted a hanging solar bulb on his desk, so we used an AI agent, parametric CAD, and a 3D printer to make a stand.

I added Windows support to a tool that exports local coding-agent history. The output is useful, but it must be treated as sensitive.

I built a prototype of Warp as an ACP client, with Warp handling the terminal UI and a local process owning the agent runtime.

An AI interface should create a table, dashboard, or task board when chat is the wrong way to inspect the work.

The same government foreword used 37.4% to 67.5% more tokens in Norwegian Bokmål than in English, despite fewer characters.

AI helps a capable builder move faster. It does not remove the judgment, QA, customer feedback, and maintenance behind a real product.

Chat works for commands. Longer agent tasks also need persistent state, task-specific views, audit history, and rollback.

Agents can inspect stale bugs, duplicate issues, old PRs, and half-finished ideas without getting bored. The decision still needs evidence.

For DocEngine, I want a shared issue-to-PR process with repository instructions, automated checks, and explicit human approval.

I use AI for uncertain classification, deterministic code for final writes, and human review when a mistake is hard to reverse.

AI can produce a convincing first version quickly. Product work starts with real integrations, edge cases, tests, operations, and ownership.

OpenClaw helped investigate and fix its own bug. The agent shortened the path to a reviewed patch, while humans kept control.

An agent should prove what it completed, report failures, and record the lesson so the same mistake does not return next week.

X exported my tweets and media, but not 1,174 Grok conversations. I built a workflow to keep that history in local Markdown.

I used AI to classify 1,388 password-manager entries. Secrets stayed local, and deterministic code applied reviewed changes by ID.

I built a local Android SMS Gateway skill so OpenClaw can send texts, receive replies, and export inbox data over Tailscale.

UA1.dev checks PDFs against PDF/UA-1 in a web interface or through a JSON API for CI, scripts, and agents.

What long-term CCM work with Gjensidige taught me about technical memory, trust, accessibility, and controlled communication.

Checking one PDF by hand proves little. Accessibility has to work across templates, production, validation, exceptions, and archives.

Long-running CCM work builds knowledge of templates, integrations, business rules, production windows, and the cost of a wrong change.

A CCM platform connects business rules, customer data, legal text, templates, delivery, and archives each time an organization sends a document.

Why I started writing about CCM, document systems, AI tools, and the engineering behind them.