A ground-truth benchmark for AI brand recommendations, built from Amazon's ESCI dataset where human raters labelled which brands are relevant to each query. Ranking brands by share of voice against agreement with those judgements shows the two barely correlate. Scoring Claude Sonnet 4.5 against the ground truth, it is right 83% of the time it names a judged brand but recovers only a quarter of the relevant set. Live dashboard and reproducible pipeline.

Why AI visibility tools that query a model API don't measure what customers actually see in a chat app. A controlled experiment across seven brands found that turning search off removed every citation and changed the leading brand, and that the same model cited roughly eleven sources per answer via API versus zero in the app. Proposes an eight-field reporting standard.

How to automate routine software access provisioning while keeping human review where it matters. The central design decision is that the language model never makes the access decision. A deterministic rules engine owns authorization, and Claude on Amazon Bedrock handles tool discovery and employee lookups. Covers why that separation matters for determinism, auditability, and prompt injection, plus the security details that separate a demo from a deployment, including cryptographic approval tokens with expiry and DynamoDB conditional updates that prevent concurrent approvals.

A walkthrough of an autonomous research agent that replaces the daily competitor tracking product managers do by hand. Covers the four-stage pipeline that separates gathering from synthesis, per-source recency normalization across community, developer, and web sources, and why a job that runs longer than an HTTP request needs an asynchronous delivery channel. Built with the Strands Agents SDK on Amazon Bedrock AgentCore, with Claude Sonnet 4 for synthesis and Amazon SES for delivery.