White-label automation builds · n8n / Claude / Twilio / Netlify
I build the systems your agency sells.
You own the client and the retainer. I build the pipeline behind it — multi-tenant, human-approved, monitored — and stay invisible. Below are two systems running in production right now, shown the way I actually see them: as run logs.
Autonomous publication with a human gate
CASE 01 · CONTENT OPS PIPELINE · LIVEThe pause is the product. Nothing publishes until the owner taps Approve on their phone. Decline is one tap too. This is what lets a client trust an autonomous system with their public site.
The problem
A niche content site (pickleball vertical) needed to update itself — weekly homepage edition, weekday research briefings, weekday marketing packs, weekly newsletter draft — without the owner writing anything, and without AI ever fabricating a fact in public.
The build
- Four scheduled n8n workflows sharing one config: weekly edition, daily research + marketing pack, newsletter draft, and site intake (signup + form webhooks).
- Claude drafts with live web search; every claim carries a source. A hard rule omits anything unverifiable — live scores are dropped unless an event is verifiably underway.
- Site reads a single content.json committed via the GitHub API; Netlify rebuilds automatically. No CMS, no database, nothing to maintain.
The part most builds skip
The site originally shipped with hardcoded demo content — a fake live score, a stale event, an invented subscriber count. I wired all four sections to honest evergreen defaults with jsdom tests proving the bridge fills them from sourced data only. An automation that can lie is a liability, not a product.
Status & honesty
Ran end-to-end in production July 2026: drafted, approved by phone, committed, deployed, announced — untouched by hand. This is my own property, built as the reference implementation of a pipeline first proven on a diaspora news site. I show it because you can inspect a real run, not a mockup.
Multi-tenant AI intake over SMS
CASE 02 · AI FRONT DESK · PRODUCTION-HARDENEDBoring on purpose. Idempotent webhooks, per-tenant auth and timezones, slot validation against a real calendar. The AI is the easy 20% — the other 80% is why it doesn't double-book or answer for the wrong shop.
The problem
Home-services shops miss a large share of inbound after-hours contact, and each missed contact is a job that goes to whoever answers next. The system needed to qualify and book — not take messages — and serve many shops from one deployment.
The build
- Multi-tenant from day one: one n8n deployment, per-shop config (auth, timezone via Luxon, calendar, voice rules), tenant resolved by inbound number.
- Idempotency via Twilio MessageSid dedupe — carrier retries and replays can never double-process.
- Slot validation: the AI can only offer appointments that actually exist. It cannot invent an opening.
- Daily digest so the owner starts each morning with everything the system did overnight.
Hardening history
Five infrastructure revisions: Basic Auth on every endpoint, replay-proof webhooks, timezone-correct scheduling per tenant, validated booking writes. Each revision came from asking "how does this break at 2am with three tenants live" — before it did.
Status & honesty
Complete and hardened; built as a go-to-market product for the HVAC vertical, not yet deployed for a paying tenant. I'd rather tell you that plainly than dress it up — the engineering is inspectable end-to-end, and it's exactly the lead-response architecture agencies are reselling today. If your agency has the client, this stack is ready to carry them.
What I bring to your builds
Patterns I ship by default
Human approval gates
n8n sendAndWait with responseType: approval — one tap on Telegram before anything touches the public. The single feature that makes clients trust autonomy.
Sourced-only content
Claude drafts carry citations; unverifiable claims are omitted, never guessed. Enforced in the prompt and tested in the bridge.
Idempotent webhooks
Provider message IDs deduped before processing. Retries, replays, and double-taps are no-ops.
Multi-tenant config
Client-specific behavior lives in one config object — onboarding client N is data entry, not development.
Exception-only monitoring
Every pipeline emits heartbeats to a monitoring dashboard I built and run. Alerts fire on failure and deploy drift. Silence means healthy.
Static-first delivery
Sites read one committed JSON via CDN. No CMS to patch, no database to babysit, hosting bill ≈ zero. Handover is a repo, not a runbook.
Working together
Built under your brand. Invisible by default.
Hourly / sprint
$85 /hour, intro rate
Workflow builds, API integrations, AI agent pipelines, rescue work on automations someone else left behind. Async-first, Mountain Time, fast turnarounds.
Fixed-scope build
$1,500–4,000 /system
A complete pipeline like either case above, delivered with docs, credentials handover, and 30 days of fixes. You quote your client whatever you like on top.
Also available: ongoing maintenance of workflows I've built, priced per system per month. Everything delivered in your client's own repos and accounts — no lock-in to me, ever.
Currently booking
Send me the workflow that's stuck, half-built, or still a sales promise.
I'll reply with how I'd build it, what it costs, and when it ships. If it's not a fit, I'll say so in the first email.
Email me the problemOr prove it right now
Press this and a real approval request lands on my phone — the same gate from Case 01, live. On a call, you'll watch me tap Approve and see the pipeline answer.