LLM Observability: Tracing, Sampling, and Catching Regressions Before Users Do

Part 5 of the AI Engineer Series. LLM observability is what classical APM evolved into for probabilistic systems. Trace structure, OpenTelemetry GenAI conventions, sampling strategies, and the regression detection that catches problems before users do.

Evals: The Single Highest-Leverage Discipline in AI Engineering

Part 4 of the AI Engineer Series. Evals are the single highest-leverage discipline in AI engineering. Deterministic assertions, LLM-as-judge with calibration, pairwise versus pointwise scoring, and the discipline that separates teams who ship from teams who hope.

Structured Outputs and Fallback Chains: How to Stop LLMs From Breaking Your Parser

Part 3 of the AI Engineer Series. Structured outputs are how you stop the model from breaking your downstream code. JSON mode, tool calling, constrained decoding, and the fallback chain pattern that holds when models misbehave.

DSPy: Stop Writing Prompts. Start Programming Them.

A practical, beginner-friendly guide to DSPy, the framework that makes your LLM apps easier to maintain. Build a working multi-agent system with LangGraph by the end.

Context Engineering: Why Retrieval Quality Decides Almost Everything About Your RAG System

Part 2 of the AI Engineer Series. Naive RAG pipelines fail at retrieval roughly 40% of the time. Chunking, hybrid search, reranking, and query rewriting determine quality more than any prompt change.

Harness Engineering: The Code Around Your LLM Calls That Keeps Production Alive

Part 1 of the AI Engineer Series. Prompt engineering gets the spotlight, but in production, most reliability comes from the system around the prompt. This is the harness layer, and most teams underbuild it.

The Complete AI Engineer Roadmap: A 13-Week Plan to Go From Curious to Production-Ready

A structured, week-by-week learning plan covering everything from harness engineering and RAG to inference internals, fine-tuning, and production guardrails — with real code, real libraries, and real tradeoffs.

The Token Discipline Stack: A 2-Week Field Report from Agentic Coding

Two weeks of Claude Code on a fresh account, 5.4M tokens, no bill-shock drama. I burned millions on a previous account before I understood how agentic coding costs work. Here is the system I rebuilt with — what earned its place, what didn't, and what I'd skip if starting over.

The Quiet Catch-Up: Where Open-Source LLMs Stand in May 2026

A year ago, picking an open-source LLM meant settling for lower quality. In May 2026, that's not true anymore. The real question now is which open model fits your work — not whether open models are good enough.

The Production Reality — Governance, Security, and ROI for RAG Systems in 2026

Part 5 of the RAG Series: The unglamorous work that keeps RAG alive in production. EU AI Act obligations, RBAC at the retrieval layer, provenance tracking, and an ROI framework that makes the business case concrete.