Notes

More technical write-ups than the topic pages: more implementation detail, fewer service headlines, more nerdy sarcasm.

• Jev-style decision models · 2026-Q4
  Fixed option sets and probabilities, not parsed JSON. Jev, AnyJev on open LMs, LLM-as-Jev, and when to route vs chat in a PoC.
• When the website is the API · 2026-Q4
  No export button: treat the register or legacy form like an integration. Crawl vs extract, public filings, polling procurement checks, terms, and drift.
• Retrieval is not extraction · 2026-Q4
  RAG finds a passage. Extraction fills named fields. Same PDFs, different pipeline, metric, and UI.
• The review queue · 2026-Q4
  A person signs off in the PoC. The clicks become labeled data for prompt work, fine-tunes, and a loop you can measure.
• OCR, layout models, and Azure Document Intelligence · 2026-Q4
  Skip OCR on digital PDFs. Tesseract/Paddle for lines, a layout model plus GLM-OCR locally, Azure when scans and handwriting justify the API.
• Finding opportunities vs fixing one task · 2026-Q4
  Open-ended “what could AI improve?” is a workshop. A named job with an owner is a PoC. Mixing them wastes the eight weeks.
• What we read on a model card before a PoC · 2026-Q4
  Exact repo id, LICENSE on the weights, intended use, this size’s I/O, chat template flags. Family blogs last.
• Local inference vs an API · 2026-Q3
  When we keep weights on a client GPU, when an EU-region API is enough, and the hybrid setups that show up most.
• How we score a 4-8 week PoC · 2026-Q2
  Real samples, a metric the business already uses, cost, and a written go/no-go. Chat demos do not count.