Rate Limits, 429s, and Retries in Production
How I handle LLM rate limits, 429s, and retries in production: backoff, idempotency, and queues so a sale day does not melt the support bot.
AI Tech
Evaluating AI Copy Before It Hits the Storefront
A practical rubric for evaluating AI product copy before it ships: claims, banned words, schema fields, and brand voice — without fake quality scores.
What a Context Window Actually Means for Your Product
What an LLM context window means for your product: long PDFs, chat history tax, silent quality drop at the edges, and why RAG beats stuffing the handbook.
Multi-Model Routing: Cheap First, Expensive Only When Needed
Multi-model routing for LLMs: classify first, send easy work to cheap models, promote failures—and keep a cost ceiling on Shopify bots and catalog pipelines.
When a Cheaper Model Is the Right Model
When a cheaper LLM is the right model: draft-then-promote with mini/Flash first, flagship only on failures, and how to measure the blend.
Structured Outputs: Stop Parsing Prose, Start Returning JSON
Structured outputs for LLMs: use JSON schemas and validators so free-text stops breaking your Shopify and automation pipelines.
Logging LLM Calls Without Leaking Customer Data
How to log LLM calls without leaking customer data: PII redaction, sampled traces, debug-enough retention, and Shopify support-bot gotchas.
Embeddings 101: Why Store Search Feels Dumb (and How to Fix It)
Embeddings 101 for Shopify search: why keyword search fails, how chunking and refresh cadence work, and why hybrid keyword + vector still wins.
Designing an AI Support Bot That Hands Off to a Human
How to design an AI support bot that hands off to a human: confidence signals, empty retrieval, Slack/email escalation, and a Shopify-ready checklist.
RAG vs Fine-Tuning: What I Actually Ship for Small Catalogs
RAG vs fine-tuning for small Shopify catalogs: why I ship retrieval for product and policy truth first, and when a fine-tune actually earns a seat.