Blog

AI agents, RAG, and integrations — notes from production.

Human-in-the-Loop AI: Where to Put the Person

Human-in-the-loop AI: the person sits before the action, after it, on the exceptions, or on a sample - four placements, each with different trade-offs.

Retrieval-Augmented Generation: The Failure Modes Nobody Demos

RAG failure modes: weak retrieval grounding a confident wrong answer, stale content, chunking that splits the answer, and losing the source citation.

How to Keep an LLM Feature's Cost Predictable as You Scale

LLM cost control at scale: route simple calls to a small model, cache the head of the distribution, cap calls per task, and budget for the peak multiplier.

Prompt Versioning: Treating the Prompt as Code

Prompt versioning means the prompt is code - in version control, pinned per decision, tested against an eval set, and rolled out and back like any other change.

What a Production-Ready AI System Needs That a Demo Doesn't

AI demo vs production: the demo shows the happy path. Production needs failure handling, eval sets, audit logging, cost controls, and a human escalation route.

How to Scope an AI Project So the Timeline Doesn't Slip

AI project scoping turns a vague ask into a buildable spec: non-functional constraints, one hard requirement, a first-release cut, and what's out of scope.

B2B and Parts Catalogs: Why Search Needs Exact Match First, Semantic Second

B2B parts catalog search is known-item first - part numbers, codes, and fitment need exact lexical match; semantic retrieval covers the descriptive queries.

Semantic Search for Fashion: Style, Occasion, and Fit Queries

Semantic search for fashion ecommerce handles occasion and style queries only if that data is in the catalog - visual similarity is a separate image problem.

Semantic Search for Grocery: Substitutions, Pack Sizes, and Dietary Filters

Semantic search for grocery handles substitution and recipe queries well - but dietary and allergen constraints belong in hard filters, not soft ranking.

What to Put in a Product Embedding: Composing the Text You Index

Product embedding text should be a clean composed field - title, key attributes, category, short description - not raw HTML or boilerplate.