What Can You Build With Claude? A Production Blueprint
I lay out a reusable production blueprint for Claude automation: pick a boring, measurable workflow, treat tool access as a permission boundary, and
Tag archive
I lay out a reusable production blueprint for Claude automation: pick a boring, measurable workflow, treat tool access as a permission boundary, and
I build LLM apps in Python with the raw provider SDK, typed Pydantic tool contracts and Postgres with pgvector for RAG, adding a framework only when I can
After 10+ years building AI automation, I explain that LLM engineering consulting is about making systems behave when nobody is watching, not picking a
Most LLM apps fail in production because nobody defined what "working" meant, not because the model is bad. I lay out the checklist I use, covering
Most "AI consultants" ship demos, not systems that survive a 3 a.m. incident with retries, evals, cost caps, and observability. I help founders and teams
I killed an autonomous agent after six weeks and replaced it with a state machine calling an LLM at four steps, quality rose, cost dropped 70%, and the 3
Proximity is usually the wrong first filter for AI automation work, most of it lives in APIs, cloud consoles, and dashboards that don't get faster because
Hiring an AI implementation consultant means finding someone who owns the full stack, evals, numbers, and handover, not a prototyper who ships demos
Most of my AI integration work is data plumbing, not prompting: I map where data actually lives, find the ugliest system (usually the real source of
I've refined a five-phase AI implementation process, scoping, POC, integration, evaluation, operations, where each phase has a hard exit criterion that
I design multi-agent workflows daily and the pattern that survives production is: write the workflow as a deterministic pipeline first, then let an LLM
I build unattended systems on four properties in order—scheduling, idempotency, observability, and graceful failure—starting with boring scaffolding