The first entry, because it's the most current demonstration of how I build. Server-rendered on Vercel, a streaming AI agent answering questions about my work in real time, contact and transcript delivery via Resend, content validated with Zod at build time. Held to a performance floor, not a target.
THIS SITE
02
Clearframe OS
13-department multi-agent operating system · Claude Code
Personal operating infrastructure for running parallel ventures and career work simultaneously. Each department runs its own Claude Code CLI agent with a scoped workspace, defined access tiers (Full / Read / Restricted / Zero), and a provider health monitoring layer that surfaces failures before workflows break.
PERSONAL INFRASTRUCTURE
03
AgencyOS
Multi-agent AI operating system · 7 specialist agents
Built for a four-person boutique real estate team: orchestrator, lead qualifier, property researcher, client communications, transaction coordinator, daily deal desk, and BD coordinator. Three root behavioral contracts — HANDOFF_PROTOCOL, CASE_STATE_SCHEMA, and an Emotional Intelligence layer applying Psychology of Money frameworks to real estate client archetypes. Five complete end-to-end case flows committed to the repo.
"Ariel didn't submit a competition entry. They submitted a product launch packaged inside a competition entry."
— judge feedback. Announced as winner, rescinded on an administrative eligibility requirement unrelated to the work.
A composite mentor character derived from studying 10+ operators — Andy Grove, Charlie Munger, Bill Campbell, Bill Walsh, George Marshall, Taiichi Ohno, Seneca, Marcus Aurelius, Naval Ravikant, Kim Scott. Delivered as two layers: a Claude Project drop-in (CHARACTER, KNOWING, behavioral contract, five signature questions, seven documented blind spots, failure stories, absolute refusals) and a native iOS voice app — hold-to-speak, on-device speech recognition, Claude API streaming and extended thinking.
A strength training tracker, no framework, single file. Guided 4-session Intro Week walks a lifter through their training maxes with no spotter needed — three work-up sets, Epley formula, auto-written to setup. Session logs persist in IndexedDB; the service worker keeps it usable offline. The 'Iron Ledger' visual system (phase-coded palette, circular timer dial, set table grid) came out of a full design audit against a mockup handoff before a line of the UI shipped.
Built for fractional executives and interim operators who need to understand a new client's landscape before the first deliverable, not after day thirty. It's the Filter stage of a five-stage engagement flywheel — Source → Filter → Develop → Execute → Calibration — with state handed off between stages via files on disk, no servers, no orchestration layer. Shipped as four platform-specific folders (Claude Code, Claude Projects, Codex, ChatGPT Projects) so the same methodology, voice, and refusals run identically regardless of which AI tool the operator already uses.
Folder-based Claude Project specialist · ICM architecture
A disciplined acquisition analyst for micro-PE operator-buyers screening small HVAC and Fire & Life Safety trades businesses ($500K–$5M revenue) — people who want to own and run the business, not flip it. Drop the folder into a Claude Project and it returns a structured PURSUE / PROCEED WITH CAUTION / PASS verdict, domain-specific deal logic, ranked risk flags, and data-gap calls before time gets spent on a bad deal. Deal memory accumulates into a calibration log, so each subsequent screen is more accurate than the last.