Agentic · 7 systems
What I build with agents.
I led the introduction of agentic coding to an enterprise engineering team, and I run Frenti — an AI-native boutique studio — on the same discipline. Agentic coding plus a spec-driven workflow lets one person ship like a five-person team.
The leverage isn't in the typing. It's in the spec, the review, and the judgment about what to ship. Every non-trivial piece starts as a markdown spec — goal, phases, acceptance criteria — and the implementation follows the contract. Generated content passes through an LLM-as-judge gate before it ships: scored against tasting notes for accuracy and voice, with anything below threshold routed to manual review.
Below, each product through the one mechanism that makes it agentic. For the mechanics in full — the spec discipline, the judge gates, the retrieval and pipeline architecture — see the engineering behind them, or the longer story in The Software House of One.
An agentic editorial pipeline — content brain, render layer, judge layer — writing sixteen destinations in two languages, scored on five weighted criteria before anything ships.
Conversational discovery over a taste graph. Every restaurant is written by an agent and LLM-as-judge–scored before it ships, with a WhatsApp loop for retention.
Editorial AI bridging restaurants, surf spots, and neighborhoods — judge-gated copy across eight verticals and 200+ pages.
An AI travel companion that parses booking emails into map-first itineraries and shared, collaborative trip workspaces.
A bilingual marketplace for one quiet peninsula — agentic editorial for listings over a deposit-based, Pix-ready flow on Postgres RLS.
The AI plays the teacher — a voice-led K-2 math and spelling tutor, built for my daughter.
The internal SEO agent watching every Frenti site — GA4 + Search Console in, content LLM-judged against a 0.75 threshold, then a separate content-editor agent drafts each fix as a reviewable diff for a human to approve. A closed loop. Sample output below.
Farol — SEO agent, sample output
Samplestatic demoLast run: 6 hours ago · Across 3 sites · Not connected to live Farol in this demo.