Builds we point to.
Each one a real system: problem, what shipped, and what it changed. Anonymised where a client needs it.
AI inbox & campaign platform
Problem: an editorial outreach business handling 1,000+ cold-outreach replies a month across 70 mailboxes by hand, on a creaking Airtable and Railway stack.
Shipped: a platform that classifies every inbound reply with Claude, drafts personalised editor responses, manages full email threading and runs a 7-state warmup across 69 accounts. Whole stack migrated to Vercel and Supabase with zero downtime and full audit logging.
AI document management system
Problem: an oilfield equipment company drowning in emailed certificates and documents, with no structure and no way to search them.
Shipped: an ingest pipeline that reads and classifies each document with GPT-4o Vision in a single call, files it into structured Drive folders, indexes it in Postgres with vector embeddings, and answers natural-language queries via chatbot. Plus automated expiry alerts.
10-vendor catalog scraper
Problem: a commercial upholstery business needing fabric catalog data from 10 vendor sites, several of them behind bot protection.
Shipped: 10 custom TypeScript scrapers (Shopify API, WooCommerce REST, HTML parsing, and headless browser with Cloudflare bypass) outputting one standard JSON schema: patterns, colorways, images, pricing and URLs. With rate limiting, retries and dedup.
AI pest-control planner
Problem: commercial greenhouse growers needing a 2-hour expert consultation to design each biological pest-control programme.
Shipped: a planner where the grower enters crop, facility and historical pest pressures, and Claude returns a deterministic week-by-week deployment schedule: which beneficial organisms, quantities and cost. In production, used by growers.
Survey-to-quote pipeline
Problem: an industrial infrastructure repair business turning field surveyor forms and reference photos into customer quotes by hand, with no way to trace a quote line back to the evidence behind it.
Shipped: an end-to-end pipeline where the survey stays the source of truth and vision AI sits on top purely as verification: photos never set prices, they cross-check what the survey claims and flag discrepancies for a human. Gemini 3.1 Pro detects objects in the photos with a crop-and-reverify pass, Claude Opus 4.8 gates in as a second verifier on low-confidence cases, and every quote line traces back to the exact source cells and photos that produced it. Hardened with HMAC-signed webhooks, SSRF and zip-slip guards, PII hygiene, and 1,400+ automated tests.