Blog

Notes on AI engineering in production

On the phone, silence is obvious. 0 missed calls, 96.7% success, Live in 2 weeks.

Oct 22, 2026 · customer support · voice AI · chatbots · AI agents

AI voice agent or chatbot for customer service: how to choose

Phone or chat? Pick the channel your customers already use. A practical comparison of latency, scope, human handoff, compliance and cost drivers.

Read

Pilots stall on decisions nobody made before the demo. 75% → 81%, 2 weeks, ~30 checks, EU cloud, −80% cost.

Oct 20, 2026 · AI in production · proof of concept · AI strategy · evaluation

Why AI projects never get past the prototype (and what the ones in production have in common)

AI pilots usually stall on decisions nobody made before the demo. The five things my production projects settled before building the model.

Read

No extracted value goes into the ERP on trust. ~30 checks per document, −84% manual work, 2,500+ documents a day.

Oct 14, 2026 · back-office automation · document AI · data extraction · OCR

Extracting data from PDFs, delivery notes and quotes with AI

Supplier PDFs, delivery notes, quotes and scans still get re-keyed by hand. What AI extraction does reliably today and how to keep its errors out of your ERP.

Read

Don't trust "done". Review before merge. Rule, Script, Gate, Hook, Review.

Oct 14, 2026 · claude-code · ai-agents · developer-workflow · hooks · code-review

The most common Claude Code mistakes (and how I removed them from my workflow)

Ten Claude Code mistakes I made or saw in my own repos, and the rule, hook or review step that now stops each one before it reaches main.

Read

Oct 14, 2026 · claude-code · ai-agents · developer-workflow · git-worktrees · engineering-process

My Claude Code workflow: one orchestrator and many parallel workers

One Claude Code session plans and reviews, never edits. Every task goes to a worker in its own leased git worktree. The setup, the scripts and what broke.

Read

Short, imperative, no history. Rules the agent can't infer, Pointers to docs and skills.

Oct 14, 2026 · claude code · ai agents · developer experience · ai engineering

CLAUDE.md done right: the instructions that actually matter (and the useless ones)

A short, imperative CLAUDE.md with rules the agent can't infer, pointers to docs and skills, and no history. Real examples from a working repo.

Read

Oct 14, 2026 · ai engineering · coding agents · developer productivity · claude code · engineering leadership

From developer to AI-directed engineer: what changes in daily work

Less typing, more specifying, guarding and reviewing: how daily work changes when coding agents write the first draft, from a tech lead who works this way.

Read

Right team, first time. 1M+ tickets a year, 400+ categories, 81% vs 75% human, 5–6 → 2–3 days.

Oct 14, 2026 · customer support · ticket triage · email · classification

AI ticket triage: how automatic routing of support tickets and emails works

An LLM reads each ticket or email and routes it to the right team. At a bank: 1M+ tickets a year, 81% routed correctly vs 75% for humans.

Read

Oct 14, 2026 · scheduling · automation · back-office · operations

Automated staff scheduling: stop building rotas by hand in spreadsheets

Automated staff scheduling: rotas that respect availability, contracts and rest rules, fewer hours lost in spreadsheets. When it pays off and when not.

Read

Score the work first. Automate second. Volume, Rules, Cost of error, History, Input format.

Oct 14, 2026 · automation · back-office · SMB

Which back-office tasks to automate first with AI

Start with tasks that are high volume, follow rules you can write down and fail cheaply: inbox triage, document data extraction, routine checks.

Read