Blog
Technical reflections on software architecture, AI-assisted development, and complex systems.
6 min readHow I built two MCP servers for a personal need, with no official API
EcoleDirecte and the French Chess Federation expose no public API. What I built to query my own school and chess data from Claude, without ever becoming a service you connect to.
→- 6 min read
Basel III/IV, Explained to a Developer→
A standard that answers a single question, how much capital a bank has to hold in reserve, computed from the risks it actually carries.
- 5 min read
Default Risk, Explained to a Developer→
PD, LGD, Expected Loss: three building blocks that answer a single question, how much does it cost, on average, to lend to someone who might never pay you back.
- 5 min read
Debugging a Production Incident with an Agent: What Changes, What Doesn't→
An agent reads a thread dump faster than a human can. That doesn't replace the discipline of evidence, or the risk it jumps to a conclusion from a single measurement.
- 3 min read
LangChain4j vs Spring AI: two ways to bring a LLM into a JVM→
Same JVM, same need, two opposite philosophies: one starts from the framework, the other from the model. A comparison for a Kotlin/Java developer.
- 3 min read
Spring AI: bringing Claude into a Kotlin/Spring Boot backend→
Spring AI promises to wire a LLM in like any other Spring bean. What that actually looks like on a Kotlin backend, past the hello-world example.
- 3 min read
Claude Code is not the Claude API: what MissionPilot taught me by combining them→
A dev tool augmented by an agent and a product that calls an LLM at runtime solve two different problems. Confusing them costs you at the design stage.
- 3 min read
Plan Mode: why I trigger it before any non-trivial task→
Discovering an approach problem after the agent has already edited ten files costs more than discovering it before. Plan Mode moves that discovery moment.
- 3 min read
MCP: connecting Claude to real tools, not longer prompts→
Describing every tool in text inside the system prompt doesn't scale. What MCP changes by standardizing tool access instead of narrating it.
- 4 min read
CLAUDE.md: the context you finally stop re-explaining every session→
Onboarding an agent on a project's conventions every session is time lost repeating the same thing. CLAUDE.md does it once, versioned, for everyone.
