4 min read
MCP Resources vs Tools vs Prompts: Stop Putting Everything in a Tool
The three primitives in the Model Context Protocol, what the model actually sees, and a rule of thumb for Dataverse, files, and internal APIs.
7 min read
Evaluating LLM Applications: Getting Past 'It Looks Good to Me'
Shipping LLM features on vibes works until a prompt tweak silently breaks ten other cases. Here is how to build evals that catch regressions before your users do.
7 min read
Getting Structured Output From an LLM Without the Heartbreak
Asking a model to 'return JSON' and parsing the result is how you get 3 a.m. pages. Tool schemas, constrained decoding, and validation turn a probabilistic text generator into a reliable API.
7 min read
Chunking Strategies for RAG: Where Retrieval Quality Is Won or Lost
Most RAG systems that retrieve bad context aren't failing at embeddings or reranking — they're failing at chunking. How you split documents quietly decides what your model can ever find.
10 min read
An MCP Server for Dynamics 365 Finance and Operations: Natural-Language Access to ERP
How to expose Dynamics 365 Finance and Operations to Claude through MCP — the architecture, the data entity choices, the auth model, and the operations that should never be one prompt away.
6 min read
Claude Code Hooks: Automating Repo Guardrails Without Pre-Commit Fatigue
Use Claude Code hooks to enforce policy at the moment actions happen — before edits, before tool calls, on session start — without relying on pre-commit hooks everyone learns to bypass.
7 min read
Building Your First MCP Server in Python: A Hands-On Walkthrough
From zero to a working Model Context Protocol server in about 100 lines — tools, resources, a local client test, and the traps that will bite you on day one.
7 min read
Prompt Caching with Claude: What It Saves, Where It Fails, and How to Set It Up in Production
A precise look at Claude's prompt caching — how the cache hit math actually works, the 5-minute and 1-hour TTL variants, and the subtle mistakes that silently disable it.
8 min read
Tool Use with Claude: Parallel Calls, Retries, and the Patterns That Survive Production
The tool-use features of the Claude API look straightforward in the docs. These are the production patterns that separate a demo from a system that handles real traffic — parallel tool calls, retry loops, error recovery, and tool_choice tactics.
7 min read
Claude Agent SDK vs Direct API: When to Reach for Each
The Claude Agent SDK is the fastest path to a capable coding-style agent. The raw messages API is the fastest path to understanding what your agent is doing. Here's when each is the right tool.
8 min read
Designing a Production-Ready RAG System: What Actually Matters
Retrieval-augmented generation demos look great. Production RAG is a different engineering problem — chunking, hybrid retrieval, reranking, evaluation, and the failure modes nobody mentions at conferences.
8 min read
Vector Databases Compared: pgvector, Qdrant, Pinecone, and When You Don't Need Any
A practical comparison of the vector databases people actually deploy in 2026 — and an honest look at when a vector database is the wrong tool for the job.
8 min read
Event-Driven Architecture for LLM Pipelines: Queues, Retries, and the Cost of Synchronous Chains
LLM-powered pipelines break in the same ways old monoliths did — when you stack synchronous calls. Queues, events, and the patterns that make a six-step AI pipeline survive a partial outage.