One platform that runs the sites, the studios, and the business.
DVNC Cloud is the infrastructure venture underneath everything else. Client sites deploy as isolated tenants on an edge runtime. Projects, proposals, invoices, and auth live in a reactive control plane. And a suite of AI studios does real production work per client: social visuals, decks and documents, directed video, video research, chat, and editorial, each billed to that client's own AI keys.
- At a glance
- 2026
- Role
- Founder, platform engineer
- Type
- Multi-tenant platform
- Year
- 2026
- Status
- Live, in production
A control plane for the whole client portfolio.
Running many client sites and many client deliverables by hand does not scale. DVNC Cloud turns both into a platform: each client is a tenant with its own isolated site runtime, and each client gets its own AI studio state, so creative and engineering output is produced, reviewed, and delivered from one dashboard.
The planes never mix. What clients see runs on Cloudflare. Business data runs on Convex. Source lives on GitHub. If one degrades, the others keep working.
- Multi-tenant
- isolated worker per client site
- 6
- AI studios and agents on one dashboard
- 1
- account pays for each AI call, no markup
Many sites, many deliverables,one pair of hands.
Hand-deploying client sites is the bottleneck that caps an agency, and AI work has the same problem: every client needs visuals, documents, video, and research, with their own brand, their own data, and their own costs. The platform problem is isolation: between tenants, between clients' AI state, and between who does the work and who pays for it.
- Per-client deploys that do not scale past a handful
- No isolation between one tenant and the next
- AI spend that blurs across clients without clean billing
- Client work that needs review before the client ever sees it
What I built,and how it came together.
The buildIsolated tenant sites, per-client AI studios, one reactive dashboard.
A Convex control plane for data, auth, and billing. A Workers for Platforms runtime that gives every client site its own isolated worker. A set of AI studios that hold each client's creative state and produce real deliverables.
- Cloudflare Workers for Platforms: one isolated worker per tenant site
- Static Astro tenant sites, editable from the dashboard, rebuilt on publish
- AI studios per client: visuals, decks and documents, video direction, video research
- Every AI call billed to exactly one client's own keys, no markup, no shared pool
- Draft-first client work: admins build with the client's material, clients see nothing until it is sent
- 01Split the planes
Client-facing on Cloudflare, business data on Convex, source on GitHub. Failure in one never takes the others down.Architecture · Tenant contract · Boundaries
- 02Build the runtime
Workers for Platforms dispatches an isolated worker per tenant site, with no tenant-side secrets, ever.Dispatch namespace · Tenant isolation · R2 storage
- 03Build the studios
Each studio is its own stateful agent per client account: brand approaches, styles, runs, and deliverables, all reactive.Visual studio · Deck studio · Video direction
- 04Wire the money law
One account pays for every AI call it triggers, through its own vaulted keys. There is no platform key anywhere in the runtime.Key vault · Per-account billing · Cost visibility
The parts that make ita platform instead of a pile of tools.
Tenant isolation by construction
Every client site is a static build served by its own worker inside a dispatch namespace. Tenants carry zero secrets, are unreachable except through the platform's dispatcher, and cannot see each other. A client editing their content triggers a rebuild of their site alone.
The AI studios
A visual studio plans and renders full social sets against each brand's own art direction. A deck studio plans documents as typed page trees and renders any layout from one renderer. A video desk fires prompt sets at generation models with per-run cost ceilings checked before anything is billed. A watch desk turns any YouTube video into a scored decision report with deep links. Each studio holds its state per client, and clients can be granted any studio as an entitlement.
One account pays, always
Every model call on the platform is billed to exactly one account's own vaulted keys, resolved at call time. There is no platform fallback key, so no client's work can ever silently spend someone else's money. Keys are probed at paste time, and admins can see what every run cost.
Draft-first client work
An admin can work for a client, with that client's voice, brand, and keys, while everything produced stays a draft the client cannot see. One explicit send delivers the finished piece to the client's content inbox. Building for someone and publishing to them are two different acts, enforced where the data lives.
The production fleet
Behind the dashboard, one machine per job shape writes articles, audits sites, plans coverage, and distributes to social for the client brands. Prompt material is data, deployed like code, and every run leaves a full timeline that can be replayed when something goes wrong.
Failure is a design input
If the control plane degrades, tenant sites keep serving; if the edge degrades, the dashboard keeps working. Long-running jobs are queued, retried with backoff, and swept by a pulse that turns orphans into honest errors instead of silent hangs.
The work
Visit dvnc.cloudClient sites as tenants,client work as a product line.
DVNC Cloud turns a one-person practice into a platform: production tenant sites served from isolated runtimes, AI deliverables produced per client with honest per-client billing, and one dashboard running all of it.
Built with
Start with a short call. Straight answer either way.
Tell me what you are trying to ship. We scope it, price it, and decide the right way in.
One letter, every Sunday.Working systems, not hot takes.
Weekly. No spam. Unsubscribe anytime.