Blog
Notes from practice
Thoughts on AI inside companies, agentic workflows, web apps, and how we work - from running projects, not slide decks.
- 9 min readAgenticSebastian Gieseler
When several coding agents work on one stack at the same time, crosstalk between them is the most expensive failure: a migration in the wrong database, a job in the wrong stack, a build that never arrives. What our days looked like before devstack, and why every worktree now gets a self-contained workspace.
- Agentic
- Open Source
- Arbeitsweise
- 8 min readOpen SourceSebastian Gieseler
Shipping almost every day means deploying into running background jobs almost every day. How we turned the deploy window into a predictable pipeline step with Croniq's maintenance mode, a configuration gate and run_on_register.
- Open Source
- DevOps
- CI/CD
Our open-source project QSOlog logs radio contacts entirely in the browser: offline-first PWA, data in IndexedDB, ADIF export. What the project shows about web applications without a backend.
- Open Source
- WebApp
- PWA
- 3 min readAgenticTyron Carlomagno
Claude Code, Codex, and Gemini CLI are impressive in demo repos. In mature codebases, the deciding factor is how tokens, reviews, and antipatterns are managed - not the model.
- Agentic
- Ways of Working
- 4 min readOpen SourceTyron Carlomagno
Parsing cron expressions is a solved problem. It gets hard when a runner disappears mid-job. What building Croniq taught us about orphaned executions, identity takeover and calendar gating.
- Open Source
- DevOps
- 4 min readAI PracticeTyron Carlomagno
OCR reads letters. Intelligent Document Processing understands documents. The difference decides whether incoming mail, case files or invoices can really be automated.
- AI Practice
- IDP
- Local AI
- 4 min readWays of WorkingTyron Carlomagno
Good software does not end at go-live. A codebase another team can take over within two weeks is a trust signal - and usually better engineering.
- Ways of Working
- 2 min readAI in PracticeTyron Carlomagno
A working RAG prototype is 30 % of the job. The decisive 70 % sit between pilot and production - and decide whether case workers actually open the tool on Monday morning.
- AI in Practice
- Local AI
The Model Context Protocol is more than an acronym in the hype cycle - but not every integration deserves its own MCP server. When the effort pays back and when a plain script will do.
- Agentic
- AI in Practice
Local does not automatically mean compliant. Between 'runs on-prem' and 'survives an audit' sit policies, a vault for personal data, and a verifiable audit log. An assessment using the open-source project Powerbrain as the example.
- AI in Practice
- Local AI
A well-written skill beats ten ad-hoc prompts. How system prompt, tools, and quality standards become a reusable team asset.
- Agentic
- AI in Practice
- Ways of Working
For internal line-of-business apps in SMBs, the PWA is almost always the smarter choice. In 2026 the last 'but you can't do that' arguments are gone.
- WebApp
- PWA
The open-source project Caveman makes coding agents like Claude Code or Gemini CLI talk in telegraph style - cutting output tokens by 65 % on average. A practitioner's take.
- AI in Practice
- Agentic
- Ways of Working