No abstract advice. Code, architecture and decisions from the production systems we worked on.
An internal AI assistant for a dental clinic: she orders lunches, compiles reports for management and talks to the HR system. What she does, how she is built and what a few weeks in operation taught us.
Custom software development pays off when off-the-shelf does not fit — but not always. Five signals you need custom, plus a build vs buy decision framework.
Model Context Protocol is a hot topic. Wiring Claude to a database, however, doesn't make a product. After several prototypes, we use a five-question methodology before a single line of MCP server code gets written — plus one uncomfortable question about admin UIs.
On the translation pipeline for The Clinic Praha we moved from the synchronous Claude API to the Batch API. 50% discount plus prompt caching gets us around 80% savings. The catch is the queue — and a few cases where batch is the wrong call.
Magistra DWH integrates data from 50+ sources — pharmacies, e-shop, HR, external APIs. Schema changes without warning, data quality varies. Contract testing, quality gates and observability pipeline as code.
A dental clinic with a two-language Laravel site needed to reach patients in six languages without losing Google rankings. Here's how we did it in three weeks — and how the redirect AI matcher covered 97.6% of old URLs.
The biggest barrier to leaving Shopify isn't building the new system — it's the fear of losing SEO rankings and breaking checkout during cutover. Lessons from MessyPlay: URL mapping, redirects, parallel runs, and what can't be skipped.
"Architecture-first" is one of those phrases that everyone agrees with and nobody defines. Here's our working definition — and the four things we refuse to do because of it.
For Magistra we're building a DWH over a network of 200+ pharmacies. Star schema vs. Data Vault, handling branch mergers, late-arriving dimensions, slowly changing dimensions — modeling decisions that don't reverse easily.