Work

Selected work, honestly labeled.

Filter by proof type. Holley Studio does not currently list external Studio client case studies here—what you see is prior professional work, products, internal systems, and demos, each tagged so the origin is clear.

Prior work / referenceInternal dogfood systemLive in production

Prior Work · Enterprise AI PlatformPrior professional work

AI Creative Operations Platform

Prior professional work — regional creative agency, 50+ employees (built before Holley Studio was founded)

Role: Lead engineer / systems builder · Status: Prior professional — completed

Live demo of the same content-generation pattern used to build the original platform — try it at /demos/creative-pipeline.

Context

Enterprise creative-ops tooling built in a prior professional role — not a Holley Studio client engagement.

The Problem

A high-volume creative agency was producing 40+ campaigns per month. Brand voice consistency was inconsistent across writers. Knowledge about past campaigns was siloed in individual brains. Content repurposing took 2–3 hours per piece.

Stack

Next.js 14 (App Router)
OpenAI GPT-4o-mini
Custom RAG knowledge base (pgvector)
n8n workflow orchestration
Custom agent orchestration system
Supabase + Vercel Blob

What Was Built

Custom RAG knowledge base indexing brand voice, messaging pillars, and past campaign learnings
14-tool AI platform: campaign builder, newsletter generator, shot list builder, content repurposer, case study wizard, media intelligence analyzer, and more
Agent orchestration system with 14 skill modules for multi-step campaign creation
monday.com integration for creative request intake → automated brief generation
Governance system with thousands of enforced security and quality rules across dozens of modules
Playwright e2e tests + CI/CD pipeline with release gate validation

Outcomes

✓Content repurposing: 2–3 hrs → under 10 min per piece in team measurement (labeled prior-professional outcome)

✓Brand-voice validation against indexed pillars before delivery — reduced ad-hoc rewrite loops

✓Campaign knowledge retrieval: email/Slack threads → seconds-scale search across indexed past work

✓Platform shipped to v7.6.0 and used in daily production at that employer

Evidence: Live pattern demo at /demos/creative-pipeline; narrative from prior role

Dogfood · Holley CreativeInternal Creative system

AI Photography Operations System

Holley Creative (holley-creative.com) — built for our own photography business

Role: Founder / builder · Status: Live — Holley Creative production

intake-batch --folder ~/exports --batch-key smith-wedding-2026

Context

Internal Holley Creative operations system — dogfooded daily.

The Problem

Managing shoots from Lightroom export to client delivery required 5+ manual tools: export preset, FTP upload, CDN URL copy-paste, Supabase insert, portfolio publish. No EXIF privacy protection, no structured intake record, no AI assistance, and no connection between uploaded files and database records after publish.

Stack

Next.js 16 (App Router)
Cloudflare R2 (CDN — 3 derivative sizes per image)
exiftool-vendored (EXIF filter + copyright injection)
sharp (WebP conversion at 400/1200/2400px)
blurhash (blur-up load placeholders)
Claude Haiku Vision (AI category + alt text at $0.05/100 images)
Supabase PostgreSQL (full intake audit trail)
n8n (batch-complete notification automation)
Zod 4 + cookie-based admin auth

What Was Built

CLI intake pipeline: Lightroom JPEG export → EXIF filter (GPS stripped, copyright injected) → 3 WebP derivatives → R2 → Supabase in one command
Three responsive derivatives per image at intake time: 400px (admin grid/mobile), 1200px (portfolio grid), 2400px (lightbox/hero) — generated in parallel
AI classification: Claude Vision suggests category and writes accessibility alt text per image, reviewed before applying
Admin review grid: blurhash blur-up placeholders, optimistic approve/reject UI, AI suggestions displayed inline
Publish flow: approved assets → portfolio item with srcset responsive images + junction table for full asset traceability
Full traceability: every image trackable from source SHA-256 hash → R2 key → published portfolio item
n8n batch-complete notification: email with review link fires when CLI finishes
Multi-service booking system: contact form, status pipeline, auto-reply, follow-up sequences

Outcomes

✓Lightroom export → published portfolio: ~45 min manual process → under 5 min automated CLI (internal timing)

✓GPS and camera serial stripped from every image before upload — no manual scrubbing step

✓Copyright embedded in every WebP derivative at the file level

✓Portfolio grid serves sized derivatives with srcset instead of a single full-resolution JPEG per cell

Evidence: Running on holley-creative.com intake/admin flows

Internal Tool · Live NowInternal Studio system

AI Readiness Lead Scoring Funnel

Holley Studio — internal tool, live at holley.studio/audit

Role: Founder / product builder · Status: Live at holley.studio/audit

The live AI Growth Audit funnel at holley.studio/audit — step 1 of the 4-step qualification form. — The actual funnel, live now at holley.studio/audit.

Context

Internal Studio lead-qualification system — not a client case study.

The Problem

Most leads from a website contact form are unqualified — wrong budget, wrong timeline, wrong fit. We needed a way to automatically qualify prospects and route them to the right service tier before a first call.

Stack

Next.js 16 (App Router)
Claude claude-haiku-4-5-20251001 (personalized findings)
Supabase (audit_requests table)
n8n (score-tiered email sequences)
Zod (form validation)
React Hook Form

What Was Built

4-step multi-screen audit form capturing 10 data points about business maturity
Deterministic scoring algorithm (0–100) weighted by revenue, team size, tools, and AI experience
4 service tiers mapped to score ranges: Starting Line → Getting Started → Building Momentum → AI-Ready
Claude-generated personalized findings — 3 specific gaps with estimated costs and recommended services
n8n automation: score-tiered follow-up email sequence based on which tier the prospect lands in
Supabase storage of all audit submissions for CRM-lite pipeline management

Outcomes

✓Prospects self-score before booking — discovery calls start with a shared readiness context

✓Tier + gap summary arrives with the lead, shortening qualification on first call

✓Personalized findings (3 estimated cost gaps) keep the funnel from feeling like a generic quiz

✓Same pattern is reusable as a lead magnet for service businesses

Evidence: Public product surface at /audit

FSL-1.1-MIT · LiveOpen source

Thesmos — AI Code Governance System

Holley Studio — built for internal use, released under FSL-1.1-MIT (fair source) for the community

Role: Founder / product builder · Status: Live product · FSL-1.1-MIT fair source

The Thesmos landing page inside the Pantheon — rule engine, ecosystems, and governance pillars. — The Pantheon's live Thesmos page — thesmos-governance on npm.

Context

Studio product / open-source release — not a client engagement.

The Problem

AI coding tools (Claude, Cursor, Copilot) generate code fast but with repeating classes of security holes and anti-patterns — SSRF via fetch(userInput), admin clients in browser bundles, JWT decode without verify. Generic ESLint rules don't catch them. Teams had no systematic way to audit AI-generated code before it shipped.

Stack

TypeScript CLI (Node.js)
Rule engine across security, AI safety, auth, Next.js, React, and more
GitHub Actions integration
VS Code extension
npm package: thesmos-governance
Vercel (live scanner UI)

What Was Built

Governance engine covering security, AI safety, auth, Next.js, React, performance, database, and related categories
CLI scanner: runs locally or in CI — exits non-zero on BLOCKER findings
GitHub Actions gate: blocks PR merges when BLOCKER rules fire
VS Code extension: highlights violations inline as you type
npm package (thesmos-governance): install in any JavaScript/TypeScript project
Live web scanner — paste code, get findings instantly

Outcomes

✓BLOCKER-severity findings can be enforced in CI before merge

✓AI-specific rules target vulnerability patterns common in LLM-generated code

✓Released under FSL-1.1-MIT (fair source) — installable in JS/TS projects

✓Live scanner available for paste-and-scan demos

Try the live scanner →

Evidence: npm package + live scanner UI