Case studies
Deep technical case studies on systems I have designed, built, or led end-to-end. The architectural decisions, the trade-offs, the failure modes, and the lessons I would carry forward.
The case-study cluster. These are the longest and most technical pieces — they sit at four to seven thousand words each and are written for an architect or senior engineering audience that wants to understand the actual decisions, not just the headline outcome.
Platforms I have built end-to-end
Meridian: building the EA platform we couldn't buy
A case study in replacing a stack of licensed enterprise SaaS with one custom-built, AI-native platform. Next.js 15 + FastAPI + PostgreSQL on Azure + Microsoft Entra ID + Google Gemini. Architected, designed and personally built end-to-end. Approximately $170K/year saved at steady state. Includes the CANVAS approval workflow as a first-class surface.
CANVAS: building the approval workflow no commercial product covers
The companion to Meridian. A Nuxt 3 + Flask + PostgreSQL + Microsoft Entra ID + Azure OpenAI workflow engine that takes a new application or vendor from intake through to contract signing as one continuous, auditable system. Modular monolith, built by one full-stack developer to a six-month MVP target. Append-only audit log, single-use SLT email tokens, AI-generated architecture memos.
Springboard: AI-native job-search platform
A solo-built Next.js + Drizzle + Supabase platform with a fleet of GitHub Actions cron workers. Monitors approximately 1,400 companies, evaluates roles against a six-block rubric, generates tailored CVs and interview prep, auto-fills applications across four interchangeable plans. Approximately fifteen thousand lines, one engineer, around four dollars a month in total LLM and infrastructure cost.
Programmes architected
Lessons from large-scale ERP transformation
Drawn from leading the architecture for a CHF 350M+ multi-region SAP S/4HANA transformation programme. Five lessons that don't appear in the vendor narrative or the typical analyst commentary.
Shorter pieces from the field
These are not full case studies, but read alongside the longer pieces they give additional texture on what working on these systems was actually like.
You can build it with AI. You can fix it with AI too. Here is the proof.
The 23-bugs story from an early Springboard iteration — what happens after you ship with AI assistance and the real bugs start arriving.
OpenClaw: too good to ignore, too risky to install — that's what changed
A specific tooling decision and the architectural reasoning around it.
Related
The architecture practice cluster (/blog/architecture-practice/) and the AI and governance cluster (/blog/ai-and-governance/) provide the conceptual framing for many of the decisions made in these case studies. The full chronological index is at /blog/.