Never Send the Image to the Fine-Tuned Model
You fine-tune the voice, not the perception. Perception is a commodity you rent; voice is the thing you cannot buy.
Notizen, geschrieben während der Arbeit.
Technische Posts, Produktentscheidungen und ehrliche Beobachtungen, geschrieben beim Bauen, nicht danach.
KI-Engineering
Wie sich KI-Funktionen verhalten, wenn echte Menschen sie nutzen, nicht in kontrollierten Demos.
What Actually Breaks When You Put a LoRA Adapter in Production
The expensive failures were not training failures. They were infrastructure failures that produced plausible output instead of an error.
Harness Engineering: The Model Is a Component, the Harness Is the Product
You don't ship a model. You ship everything around it.
Loop Engineering: Designing the Cycle an Agent Runs In
An agent is a while loop. Everything hard is in the exit condition.
Tool Design for LLM Agents: APIs for a Consumer That Can't Read the Docs
Your tool's description is its documentation, its type system, and its onboarding.
Why I Won't Sell "Unlimited AI"
Every AI call has a marginal cost. 'Unlimited' on a fixed price means one of three things: weaker models, hidden limits later, or burning cash.
Your MCP Server Is Only as Good as Its Tool Descriptions
An MCP tool is not useful because it exists. It is useful when the agent understands when, why, and how to call it.
How I Would Build an Agent-Readable Personal Hub From Scratch
Start with the data model, then expose the same truth through pages, APIs, MCP tools, and examples.
API vs MCP: When Should You Use Each?
Use the interface that matches who's reading, humans, software, or agents.
Why Local-First Time Tracking Protects Your Privacy
Time tracking data is behavior data, and behavior data deserves local-first defaults.
Building a Public API for My Personal Hub (and AI Agents)
If an AI system is going to summarize your work, give it structured context instead of making it scrape and guess.
AI Agent Architecture: Patterns That Actually Work in Production
Agents fail at the seams, not the steps.
Prompt Engineering for Production: Moving Beyond Demos
Prompts that work in demos fail in production because demos don't have variance.
AI builders: chase primitives, not model size
Models impress. Systems ship.