Foothold America

The pattern

Three production AI systems for a B2B services company, all built on the same architectural pattern. The pattern: Make orchestration into Supabase Edge Functions, structured-output classification under JSON schemas, Postgres with row-level security, a React/TypeScript admin queue on Vercel, a human-review gate, and a verdict-feedback loop that recalibrates the AI against reviewer outcomes. The systems are different products. The pattern is the muscle.

1. Compliance Feed

A regulatory-intelligence pipeline that ingests thirteen-plus sources daily — DOL, EEOC, IRS, USCIS, OSHA, Federal Register (CFPB, DHS, Wage & Hour), FDIC, Federal Reserve, OCC, HR Dive — runs each through a structured-output classifier with a stricter v2.2 relevance gate that distinguishes operationally actionable items from feature stories and single-employer enforcement actions, then rewrites the surviving articles in Foothold's house voice at 450–750 words with a NYT-style structure (lede, change, context, implications, timing, what-to-watch). Service-line tagging across six product lines. Admin queue with Knowledge-Hub-grade typography, navy-gradient hero, AI-badge popover, rate and share affordances. Mid-build handoff to Foothold's external Bubble engineer for the client-portal integration. Shipped in a single calendar week.

2. Chat support integration

An AI layer on Foothold's customer-support surface, designed to answer first-tier inquiries against the same underlying knowledge base the company's human team draws from, with escalation hand-off to live staff when the model is uncertain or the question crosses defined risk thresholds. Built to reduce response latency on the most common inquiry shapes without surrendering the editorial standard the company has spent years building.

3. Incoming-growth automation pipeline

A lead-generation sequencer that runs the same architectural pattern as the Compliance Feed: source-driven ingestion, structured-output classification, ICP scoring against a versioned configuration, human-verdict capture, and a feedback flywheel that proposes recalibrated weights based on what reviewers approved and rejected. Seeded with the active Ireland Series A/B cohort and three Tier-1 web-search sources. Six-table Supabase schema (sources, prospect signals, prospects, ICP config, source runs, weight proposals) with two SECURITY DEFINER RPCs for the verdict-flywheel mechanics. Sandbox build, day-zero, on the same architectural rails as Compliance Feed.

Why this matters

FAI is the production case study for what an AI integration lead actually does inside a B2B services company. The Compliance Feed runs every day, generates real editorial output, and ships to a real client portal. The chat support integration reduces latency on real customer inquiries. The growth pipeline is the same flywheel architecture applied to a different surface. None of this is demo work. All of it ships to operating environments where the editorial standard and the legal stakes are non-trivial. The week of May 2026 in particular was: idea on Monday, scoped on Tuesday, schema applied on Wednesday, AI iterated three times in production by Thursday afternoon, live by Friday. That cadence is the thing. The pattern is the thing. Everything else is a different surface area for the same muscle.

Next in AI Experiments → Machine's Learning