// Journal
Field notes from the build.
Playbooks, teardowns, and points of view on shipping production AI — LLM pipelines, RAG, agents, evaluation, and the operational work that keeps them honest.
-
Case study 9 min readHow AI customer retention actually works
A retention engine is a lifecycle agent, not a recommender. Lessons from rebuilding Ferry's engagement stack: segmentation, NBO agents, guardrails.
Read -
How-to 9 min readHow recommendation engines actually scale (lessons from Compass)
Two-stage retrieval, feature weighting, and the latency crossover at 10M items. The architecture we use for recommendation engines that scale beyond a demo.
Read -
POV 9 min readHow AI proctoring actually works (and why most implementations fail)
AI proctoring isn't a model — it's a multi-signal pipeline with a trust problem. The mechanics, three failure modes that kill naive builds, and what works.
Read -
How-to 9 min readWhy your AI agent fails silently (and how to stop it)
The most dangerous AI agent failures aren't crashes — they're silent wrong answers. A breakdown of failure modes and the observability patterns that catch them.
Read -
Comparison 8 min readRAG vs Fine-Tuning: When to Use Which
A first-principles breakdown of when retrieval-augmented generation earns its complexity and when a fine-tuned model is the cleaner bet. No fluff.
Read -
How-to 9 min readWhy your RAG is retrieving the wrong documents
The failure point in most production RAG systems isn't the LLM — it's retrieval. A first-principles breakdown of vector search limitations and how to fix them.
Read -
POV 8 min readHarness engineering: build the test rig before the model
Most AI builds fail because the model ships before the harness that measures it. How we engineer evals, guardrails, and a repeatable delivery rig.
Read