Ai on Darek's Blog http://blog.dwornikowski.com/tags/ai/ Recent content in Ai on Darek's Blog Darek's Blog http://blog.dwornikowski.com/openg.png http://blog.dwornikowski.com/openg.png Hugo -- 0.147.2 en-us Sat, 04 Jul 2026 10:00:00 +0200 How MCP Works: What It Exposes, How It Talks, and How It Got Here http://blog.dwornikowski.com/posts/how-mcp-works/ Sat, 04 Jul 2026 10:00:00 +0200 http://blog.dwornikowski.com/posts/how-mcp-works/ A ground-up explainer of the Model Context Protocol: the host-client-server model, the primitives servers and clients expose to each other (tools, resources, prompts, sampling, elicitation), the transports, and the protocol's evolution from the 2024 launch to the 2026 release candidate. MCP Auth, Explained: Every Method, Direct and Through a Proxy http://blog.dwornikowski.com/posts/mcp-auth-explained/ Sat, 04 Jul 2026 09:00:00 +0200 http://blog.dwornikowski.com/posts/mcp-auth-explained/ A field guide to authentication and authorization in the Model Context Protocol: OAuth 2.1 with PKCE, dynamic client registration, client ID metadata documents, machine-to-machine flows, and what changes when an MCP gateway sits in the middle — token passthrough, token exchange, on-behalf-of chains, and ID-JAG. Prove It: Two Gates, No Mocks http://blog.dwornikowski.com/posts/prove-it-dual-testing/ Tue, 09 Jun 2026 09:00:00 +0200 http://blog.dwornikowski.com/posts/prove-it-dual-testing/ A small prompt habit that changes how a coding agent verifies its own work: stop asking for passing tests, ask it to prove the feature works — once locally against a real Tilt stack with no mocks, then again on staging against real data and integrations. The Sixth Layer: When a Local Model Beats the LLM That Trained It http://blog.dwornikowski.com/posts/local-model-llm-classifier/ Thu, 07 May 2026 12:00:00 +0200 http://blog.dwornikowski.com/posts/local-model-llm-classifier/ A follow-up to my LLM cost-optimization post: how I replaced ~55% of the remaining LLM calls in a product classifier with a small local model trained on the LLM's own labels — and why the local model ended up more accurate than the labels it learned from. How I Refused to Pay €80 for a Broken Plastic Tube and Asked Claude to CAD It Instead http://blog.dwornikowski.com/posts/grohe-dispenser-ai-print/ Wed, 06 May 2026 00:00:00 +0000 http://blog.dwornikowski.com/posts/grohe-dispenser-ai-print/ A small piece of plastic in our kitchen soap dispenser snapped. The official Grohe replacement costs €80. I had a 3D printer and a Claude Cowork session — five iterations later, I had a working part. From $200 to $30: Five Layers of LLM Cost Optimization http://blog.dwornikowski.com/posts/cutting-llm-costs-token-optimization/ Fri, 24 Apr 2026 12:00:00 +0200 http://blog.dwornikowski.com/posts/cutting-llm-costs-token-optimization/ A walkthrough of how I cut an LLM-based product classifier's token usage by ~92% through compression, two-stage prompting, exact-match lookups, similarity caching, and batching.