From an idea to a working application.
Making complex execution understandable.
An agent’s work is more than a reply. It edits files, runs commands, calls tools and deploys. The interface has to show that activity as it happens and keep it available long after.
An interface that stays fast while agents work
Server rendering, streaming and Suspense, client and server caching, virtualized lists and Web Workers keep the workspace responsive under load. French and English are served from the server.
- Next.js
- React
- TanStack Virtual
- next-intl
Every step, kept
Agents run server-side through Convex Agent and durable Actions. Deltas and tool activity are written to persistent threads and pushed to the browser over reactive subscriptions, so nothing is lost when a tab closes.
- Convex Agent
- Convex Actions
- WebSocket
A choice of models
Choose from leading model families: OpenAI GPT, Anthropic Claude, Grok, Mistral, DeepSeek, GLM and more. Multimodal routing runs through Vercel AI Gateway or OpenCode Go. Each run records its model source, and personal credentials stay encrypted.
- Vercel AI SDK
- Vercel AI Gateway
- OpenCode SDK
Giving agents a reliable environment.
Agents act on real files and real commands. Isolation, interruption and recovery are part of the product, not an afterthought.
Isolated workspaces
Each session runs in a Vercel Sandbox with capability-scoped tools: files and diffs, bounded terminal commands, web browsing, environment variables, diagnostics and deployment. I/O and timeouts are bounded, mutations are guarded and secrets are redacted.
- Vercel Sandbox
Snapshots you can come back to
The sandbox is the source of truth. Snapshots are taken around AI responses, editor writes, suspends and deploys, recorded in Convex and stored as tarballs in S3 for resume, recovery and version history.
- Convex
- AWS S3
Work that finishes, even when interrupted
Orchestration across sandboxes and Convex relies on runtime leases with heartbeats, deadlines, cancellation and state machines. Idempotent reconciliation jobs keep every run, action and fallback traceable.
- Convex Actions
- Vercel Sandbox
Connecting creation to deployment.
A generated app is only useful once it can be shared, shipped and paid for. I owned everything around generation.
Deploy in one motion
Frontends and custom domains go through Vercel for Platforms. A project-scoped backend is provisioned, connected and deployed through the Convex Platform API, with environment sync, lifecycle tracking and rollback states.
- Vercel for Platforms
- Convex Platform
Sharing with the right people
Real-time collaboration with role-based memberships, permission-aware member management, hashed invitations that expire or can be revoked, and ownership transfer.
- Better Auth
- Convex
Billing that can be trusted
Usage-based Stripe billing on an append-only credit ledger: atomic reserve, debit and release, per-generation cost reconciliation and idempotent checkout webhooks. The Stripe catalog is synced from code.
- Stripe
Critical flows under test
Vitest unit and integration tests, plus Playwright end-to-end checks on the flows that matter most: project creation, deployment and billing.
- Vitest
- convex-test
- Playwright
What I delivered.
A complete path from a prompt to a deployed application.
- Persistent workspaces that resume and recover
- A live, durable record of every agent step
- Deployment of generated apps and their backends
- Collaboration with roles and revocable invitations
- Usage-based billing on an auditable ledger
These describe the capabilities built. No usage or revenue figures are published here.
- Interface, rendering and typed contracts
- Next.jsReactTypeScript
- Reactive data and APIs
- ConvexoRPCZod
- Isolated execution and history
- Vercel SandboxAWS S3
- Billing and user-flow verification
- StripeVitestPlaywright
Scope documented in Joel Tan’s CV, pages 1–2. Abstract illustrations of the product and its architecture, informed by IteruCode screenshots and implementation, September 2026.