"How do I build an MCP server?" is the wrong question. SDKs exist. An LLM can scaffold one in seconds. The real question is: What role should an MCP server play inside a system that must not lie? The scarce skill now is designing the system around MCP:…
Page 2 of 6 (Total items: 57)
LLMs (Large Language Models) are being used as sensors. That is a category error: using a probabilistic synthesizer where a deterministic boundary device is required. This isn't developers' fault. Most mainstream examples and tutorials lead with the simplest…
Status: In development as part of lucidRAG. Source: github.com/scottgal/lucidrag Where this fits: VideoSummarizer is the orchestrator of the lucidRAG family, combining five pipelines into a unified video analysis engine: DocSummarizer - Documents (entity…
Status: AudioSummarizer.Core is currently in development as part of lucidRAG, a forthcoming mostlylucid product for multi-modal RAG. The implementation is complete and working—this article documents the architecture and design. CLI integration will be…
If you're brand new to RAG, start with RAG Explained and RAG Architecture. This post is for the point where you've built a RAG pipeline that mostly works… and now you're paying for it in cost, fragility, and "why did it answer that?" debugging sessions. What…
Part 4: Image Intelligence introduced the ImageSummarizer wave architecture and the broader patterns. This article deep-dives into the OCR subsystem—three tiers of text extraction, intelligent routing, and the filmstrip optimization that achieves 30× token…
Parts 1-3 described Constrained Fuzziness as an abstract pattern. This article applies those patterns to a working image analysis pipeline that demonstrates the principles in action. CLI Tool - Command-line interface for image analysis Core Library - The…
One of the easiest ways to make an AI system fail is to let it remember too much. Most "context handling" approaches boil down to one of two mistakes: Stuff everything into the prompt and hope the model figures out what matters Repeatedly re-summarise…
Part 1 described a pipeline: Substrate, Proposer, Constrainer. One probabilistic component, bounded by deterministic truth. But what happens when you have multiple proposers? Multiple LLMs? Vision models talking to text models? Heuristics feeding into neural…
The Pattern That Kept Emerging It started with my blog's translation tool. The cache kept only 5 tasks per user, with 6-hour absolute and 1-hour sliding expiration. Sixth task? Oldest evicted. After an hour of inactivity, everything expired. Most developers…
Page 2 of 6 (Total items: 57)