House of
AI pipelines, editorial standards.
Campaign
imagery.
Speculative campaign for Salt & Stone's Lily & Yuzu Signature Scent Duo. The concept traces the ingredients to their origin — yuzu orchards in Japan's mountain valleys and wild lilies on coastal dunes. Four images, one session, from brief through art direction to final output.
AI-generated imagery / Concept campaign, not client work
Systems I
have built.
Problem
Curating a luxury directory of alcohol-free travel experiences manually doesn't scale. Every venue needs discovery, verification, scoring, and editorial review before publishing.
System
Automated discovery pipeline. Google Places API finds candidates across 15 search queries per city. Scrapers pull website and review data. Claude scores each venue on a 1-5 scale against brand criteria. Flagged venues go to a human editorial review dashboard. Approved venues publish to the directory via Supabase with ISR.
Output
73 venues in London scored and editorially reviewed so far. Directory with search, filtering, and maps. AI trip planner responding in a defined brand voice. 26 analytics events wired across the full user journey.
Impact
Still in progress. One person, one city, 73 fully audited venues and counting. The pipeline handles discovery through scoring. The human handles the editorial call. Built to scale city by city.
Stack
73
Venues scored and reviewed
26
Analytics events tracked
Problem
Three brands need consistent weekly content. SEO analysis, social drafts, competitor monitoring, Reddit engagement. One person can't run the full marketing cycle manually across three brands every week.
System
Python agent orchestrated via GitHub Actions. Pulls Search Console and GA4 data, identifies content gaps, generates drafts via Claude, monitors Reddit for relevant threads, tracks competitor homepage changes via diff detection. All output goes to Notion for human approval before publishing.
Output
End-to-end marketing automation system: SEO briefs, social drafts, competitor intelligence reports, and engagement opportunities. All surfaced in Notion with human-in-the-loop review.
Impact
An architecture for solo operators running multiple brands. The system handles research and drafting. The human makes the creative call. Designed for three brands, one operator.
Stack
3 brands
One automated pipeline
6 sources
GA4, Search Console, Reddit, competitors, Claude, Notion
Problem
A luxury golf company was running their sales pipeline, contacts, and financials across Notion pages. It worked until it didn't — no pipeline visibility, no inquiry capture, no real reporting.
System
Custom CRM with dashboard KPIs, Kanban pipeline, account and contact management, website inquiry capture via webhook, and financial tracking. Built to replace a Notion workspace, not replicate enterprise software.
Output
Production CRM used daily by a three-person team. Dashboard, pipeline views, contact management, website-to-pipeline inquiry flow, invoice tracking.
Impact
Built in a weekend. Replaced a patched-together Notion setup with a purpose-built tool that the team actually uses every day.
Stack
3 person
Team using it daily
1 weekend
From zero to production
Problem
The Diary of a CEO has 815+ episodes. Finding a specific expert insight means scrubbing through hours of video. There's no way to search what was actually said.
System
Seven-step RAG pipeline. iTunes API pulls episode metadata. YouTube captions extracted via yt-dlp. Claude Haiku cleans raw transcript files. Content chunked into ~300-word segments with sentence-boundary awareness and overlap. Voyage AI generates 1024-dimensional embeddings stored in Pinecone serverless. Flask frontend for natural-language search. Topic categorisation across nine categories.
Output
Semantic search engine returning timestamped, quotable moments from specific episodes. Search by concept, not keyword — ask a question, get the exact moment an expert answered it.
Impact
30 episodes indexed so far, producing 2,500+ searchable moments. Pipeline is modular — each step runs independently, designed to scale to the full 815-episode archive.
Stack
2,500 +
Searchable moments indexed
30 of 815
Episodes processed so far
Systems I
have built.
- AI Pipeline Design
- Rapid Prototyping
- Creative Automation
- Brand Voice Systems
- Human-in-the-Loop Review
- Photography / Editing
- Next.js / TypeScript
- Claude API / Vercel AI SDK
- Python / GitHub Actions
- Postgres / Supabase / Neon
- PostHog / GA4 / Search Console
- Figma / Lightroom