https://www.carrotcakeai.com/insights Practical thinking from Daniel Rosehill on custom models, AI agents, MCP and AI-native architecture. en Thu, 29 Oct 2026 00:00:00 GMT https://www.carrotcakeai.com/insights/the-case-for-bespoke-classifiers https://www.carrotcakeai.com/insights/the-case-for-bespoke-classifiers Thu, 29 Oct 2026 00:00:00 GMT Daniel Rosehill Models & Classifiers classifiers small-models structured-output fine-tuning crm inventory Frontier models return valid JSON every time. A small classifier trained on your own categories still wins on consistency, cost and ownership. Here's why. Text cleanup: Claude Opus 5.5 (AI). https://www.carrotcakeai.com/insights/what-is-webmcp https://www.carrotcakeai.com/insights/what-is-webmcp Thu, 15 Oct 2026 00:00:00 GMT Daniel Rosehill Agents & MCP webmcp mcp agentic-web standards WebMCP is a proposed web standard that lets websites expose structured tools to AI agents directly through the browser. Here's how it works and where it stands. https://www.carrotcakeai.com/insights/webmcp-preparedness https://www.carrotcakeai.com/insights/webmcp-preparedness Thu, 15 Oct 2026 00:00:00 GMT Daniel Rosehill Agents & MCP webmcp mcp agentic-web e-commerce How website operators, especially in e-commerce, can prepare for AI agents that interact through structured tools instead of screen scraping. https://www.carrotcakeai.com/insights/prudent-in-house-agent-skill-development https://www.carrotcakeai.com/insights/prudent-in-house-agent-skill-development Mon, 27 Apr 2026 14:00:00 GMT Daniel Rosehill Agents & MCP agents mcp skills security consulting best-practices Agent skills are quietly becoming one of the most useful primitives in agentic AI — and one of the easier ways to leak internal data into places it shouldn't go. A few practical disciplines I've adopted when building skills for client teams. https://www.carrotcakeai.com/insights/purchasing-decisions-claude-code-workspace-pattern https://www.carrotcakeai.com/insights/purchasing-decisions-claude-code-workspace-pattern Fri, 17 Apr 2026 15:30:00 GMT Daniel Rosehill Working Practice claude-code workspaces shopping purchasing mem0 israel templates A pattern I've settled into for buying things thoughtfully: one repo per purchase, standing preferences in Mem0, an Israel-specific plugin that handles the local retail quirks, and two workspace templates that split product merit from sourcing. https://www.carrotcakeai.com/insights/audio-multimodal-llm-bitrate-eval https://www.carrotcakeai.com/insights/audio-multimodal-llm-bitrate-eval Fri, 17 Apr 2026 12:00:00 GMT Daniel Rosehill Evaluations audio dictation evals gemini gpt-audio openrouter transcription voxtral An eval across 12 audio-capable models on OpenRouter, 5 MP3 bitrates, 4 dictation samples. The default upload quality in most dictation apps is 2–4× overprovisioned — and the more interesting finding had nothing to do with bitrate. https://www.carrotcakeai.com/insights/api-or-mcp-when-the-cli-just-works-better https://www.carrotcakeai.com/insights/api-or-mcp-when-the-cli-just-works-better Fri, 10 Apr 2026 13:00:00 GMT Claude Code Agents & MCP mcp api cli agents architecture developer-experience The MCP protocol promises unified tool access for AI agents, but sometimes a plain API call or CLI command is more reliable. What does that tell us about MCP's maturity? https://www.carrotcakeai.com/insights/personalised-ai-proactive-interview-based-context-generation https://www.carrotcakeai.com/insights/personalised-ai-proactive-interview-based-context-generation Wed, 25 Mar 2026 23:00:00 GMT Daniel Rosehill Architecture open-source personal-ai rag architecture An architecture proposal for generating personalised AI context through deliberate interview-driven workflows rather than passive data collection. https://www.carrotcakeai.com/insights/bias-censorship-eval-tests-llms https://www.carrotcakeai.com/insights/bias-censorship-eval-tests-llms Wed, 25 Mar 2026 21:00:00 GMT Daniel Rosehill Evaluations bias censorship open-source evals Test prompts and notes exploring cultural bias, censorship, and worldview differences across Western and Eastern large language models. https://www.carrotcakeai.com/insights/vibe-code-and-learn-subagent-model https://www.carrotcakeai.com/insights/vibe-code-and-learn-subagent-model Wed, 25 Mar 2026 21:00:00 GMT Daniel Rosehill Working Practice open-source vibe-coding agents education Planning notes for a multi-subagent system that turns AI code generation into structured learning opportunities, not just output. https://www.carrotcakeai.com/insights/on-mcp-and-consolidation-current-pain-points-solutions https://www.carrotcakeai.com/insights/on-mcp-and-consolidation-current-pain-points-solutions Wed, 25 Mar 2026 18:00:00 GMT Daniel Rosehill Agents & MCP open-source mcp resource-lists Notes and resources on the pain points of MCP adoption and why consolidation through gateways and connectors matters. https://www.carrotcakeai.com/insights/mcp-gateway-model-notes-better-architecture https://www.carrotcakeai.com/insights/mcp-gateway-model-notes-better-architecture Wed, 25 Mar 2026 18:00:00 GMT Daniel Rosehill Architecture mcp architecture stacks Planning notes and sketches for an MCP gateway architecture that aggregates servers into LAN and WAN gateways. https://www.carrotcakeai.com/insights/fine-tuning-whisper-on-modal-serverless-gpus https://www.carrotcakeai.com/insights/fine-tuning-whisper-on-modal-serverless-gpus Wed, 25 Mar 2026 17:00:00 GMT Daniel Rosehill Models & Classifiers models open-source whisper speech-recognition fine-tuning modal gpu A script for fine-tuning OpenAI's Whisper speech recognition models using Modal's serverless GPU infrastructure. https://www.carrotcakeai.com/insights/admin-mcp https://www.carrotcakeai.com/insights/admin-mcp Mon, 16 Mar 2026 00:00:00 GMT Daniel Rosehill Agents & MCP mcp admin-mcp architecture backend An Admin MCP is a Model Context Protocol server that gives AI agents structured, privileged access to your backend systems. What that means and why it matters.