Retested Week 18's Memory hypothesis against real vendor compatibility claims, found the rubric couldn't distinguish well-formed rulings from correct ones and cold-start agents were already consistent without Memory, and killed the Managed Agents approach.
Building AI Products. Building With AI.
Welcome to my laboratory! Each week, I build conversational interfaces, RAG pipelines, agentic systems, evaluation frameworks and other goodies.
Latest entry · August 4, 2026 · 10 min read
Week 20 Journal: Marketing Claim Governance Service, Part 5 — Discovery
Built a front door for agent discovery — a claims demonstration page, a machine-readable manifest, and a read-only registry endpoint — and verified that three classes of browsing agent could traverse from page to manifest to substantiation status.
Read entry →Journal
Tested whether a Managed Agent with cross-session Memory produces more consistent claim rulings than a fresh-context agent, fixed several validity bugs in the process, and found the hypothesis refuted.
Built a multi-agent adversarial substantiation workflow that pressure-tests marketing claims through parallel evidence gathering and an attack-rebuttal loop, then measured the cost delta between plan-in-context and plan-in-script orchestration.
Shipped
Claims Desk
liveA deployed marketing-claim governance service: a registry that answers deterministic questions about a claim, an adversarial workflow that argues the judgment calls, and a public record an agent can fetch instead of taking vendor prose at face value.
View project →SAFe Feature Spec System
liveSix-agent pipeline that classifies, drafts, generates, reviews, improves, and polishes SAFe feature specifications with a 100-point rubric scoring system.
View project →Dino: AI Vegas Dining Concierge
liveAI-powered dining concierge for Las Vegas with Rat Pack-era personality, real restaurant discovery via Google Maps, mock booking with a real API contract, and Google Calendar deep links with insider tips.
View project →