Running a Software Factory on a Self-Hosted AI Coding Agent Platform
A field report from week one on a self-hosted AI coding agent platform: 13 agents, 81 issues, the real token-cost fight, and why humans still review.
Arnaud Pouppeville Architect’s Notebook
Drafts on systems, data, and the AI software factory.
Latest posts
Rendering Agent Plans Into Something I Can Actually Review
Claude Code plan review falls apart in chat scrollback. Why I render every implementation plan as a browser page I can navigate, mark, and annotate.
Terraform Infrastructure Repo: What Belongs (and What Doesn't)
A practical guide to Terraform infrastructure repo structure: what belongs, how to split state by blast radius, and what to keep out.
Claude Code Slash Commands as an Interaction Vocabulary
Claude Code custom slash commands as a vocabulary: named modes with output contracts make agent output reproducible.
Turning Architecture Into a Prompt Pipeline
How to build a reproducible architecture pipeline with Claude Code slash commands, hard gates, and input chaining. Step-by-step anatomy of one command.
Architecture Artifacts in Order: The Missing Handoff Between Discovery, Design, and Delivery
How to chain event storming, ADRs, C4, and arc42 into a repeatable pipeline where each artifact feeds the next. A practitioner walkthrough with diagrams.
Why Your SQL Server View Is Slow (And How an Inline TVF Fixes It)
SQL Server views can hide an expensive GROUP BY boundary. Here's the diagnosis and how an inline TVF brought a 3-minute query to 4 seconds.
My Agent Can Say No: Spec-Driven Agent Code Review with OpenSpec
How OpenSpec CLI enables spec-driven agent code review: your agent can decline suggestions that conflict with documented architecture conventions.
The Blast Radius Problem: Running Your Coding Agent in YOLO Mode
Running your coding agent in bypass mode on your laptop is the mistake. A VM with Tailscale and default-deny egress is how you shrink the blast radius.
Production ML Without a Data Science Team
How a martech company shipped production ML with no data science team. Silver layer, AutoML, delta inference, and matchback validation to prove it works.
Natural-Language Search Over Image Libraries: RAG Architecture
How to make thousands of images searchable in plain language without CLIP: RAG architecture using vision at ingest, hybrid search, and custom retrieval glue.
KEDA Didn't Just Save Us Money. It Saved Our Jobs From Themselves.
How we stopped AKS PDF workers from OOM-crashing using KEDA scale to zero, a dedicated high-memory node pool, and a Postgres queue trigger.
Solving AI Sprawl: Using Git Worktrees and ADRs to Govern Parallel Agents
How a five-phase AI-assisted SDLC pipeline (SpecKit, Git worktrees, /done, Claude Poll, /fix-pr) enforces architectural discipline at speed without the AI sprawl.
Two Kinds of Human-in-the-Loop — And Why LangGraph Needs Both
Two kinds of HITL interrupts in LangGraph: information gaps (stuck, need data) vs authority gaps (stopped, need approval). Design for both; use interrupt(), never END as pause.
Agentic AI Governance: It Fails at System Assumptions
Agentic AI governance for mission-critical software: a three-tier model, ADR boundaries, and why AI fails at system assumptions, not syntax.