Tim FinchProducts, systems, AI
HomeProjectsWritingWork with meAbout

Open to product, platform, and AI work.

I build evidence-backed AI products and the systems needed to operate them responsibly, usually where product decisions and real workflows meet.

GitHub ↗LinkedIn ↗Devpost ↗London / Remote

In this case study

  1. 01 Overview
  2. 02 Why it exists
  3. 03 How it works
  4. 04 Built with
  5. 05 Focus
  6. 06 Architecture and AI workflow
  7. 07 Contracts, safety and reliability
  8. 08 Ownership and tenancy
  9. 09 Evaluation and deployment
  10. 10 Current maturity
←Back to work
Audience
Strategy
Repurpose
Audience signals01
Content gaps02
Growth experiments03

AI Products

CreatorOS

A strategy-first AI system for creators who are tired of generic output.

2026 / AI Products / Next.js / TypeScript / Tailwind CSS / shadcn/ui / OpenAI / Zod / Vitest

Quick read

A clear breakdown of the product, the reasoning behind it, and the implementation choices that matter.

Year
2026
Category
AI Products
Status
Active prototype

Overview

CreatorOS is a strategy-first AI product for creators and internet brands. It turns source material into a Creator Growth Pack through a sequential workflow: audience intelligence, content strategy, and strategic repurposing. The product keeps those responsibilities separate so the output can be reviewed as a system rather than accepted as one opaque model response.

Why it exists

Creators and internet brands spend too much time deciding what to make, how to adapt source material, and which content gaps are actually worth chasing.

How it works

CreatorOS breaks the work into typed, structured stages. A validated request moves through isolated agent modules, each stage passes a schema-shaped result to the next, and the final Growth Pack is composed from audience insights, content gaps, hooks, titles, repurposed content, and experiments.

Built with

Next.js / TypeScript / Tailwind CSS / shadcn/ui / OpenAI / Zod / Vitest

What I focused on

  • Designed a sequential multi-agent workflow from transcript input to a Creator Growth Pack.
  • Kept audience intelligence, strategy, and repurposing in separate modules with explicit hand-offs.
  • Used Zod contracts, bounded request parsing, and schema-hardening tests around the generation path.
  • Added production-aware Supabase adapters while retaining a deterministic local mock path for development.

Architecture and AI workflow

source material + creator context
              |
       Zod input contract
              |
  Audience Intelligence Agent
              |
   Content Strategy Agent
              |
Strategic Repurposing Agent
              |
     Creator Growth Pack

Next.js API routes -> orchestration -> isolated agent modules
Supabase adapters -> profiles, sources, experiments, snapshots
The documented flow is sequential because each stage depends on the structured output before it.

The architecture keeps the API route small and puts prompts, agent runners, orchestration, OpenAI handling, types, and storage behind explicit modules. That leaves a clean boundary for a future service split without making the MVP a distributed system prematurely.

Contracts, safety and reliability

  • Generation validates transcript, niche, audience, and platform before orchestration; request bodies are capped at 100,000 bytes and the route has a 60-second runtime ceiling.
  • The generation endpoint returns no-store responses because inputs and outputs can contain creator material, and provider or deadline failures are mapped to safe 502 or 504 responses with a request ID.
  • Creator-provided fields are treated as untrusted reference data before entering agent prompts, so embedded instructions are not treated as system direction.
  • The generation limiter allows ten requests per ten minutes per signed-in user in the shared Supabase path; production returns 503 when that limiter is unavailable instead of silently falling back to process-local state.

Ownership and tenancy

Workspace routes resolve the authenticated viewer before reading or writing data. The repository tests that an experiment cannot be created from a source owned by another user, while the production storage adapter refuses workspace operations when the required Supabase persistence configuration is missing.

  • The Supabase schema includes ownership foreign keys, valid-status constraints, non-negative metric checks, and duplicate experiment/date protection.
  • Login requests and generation requests have separate rate-limit paths, with shared Postgres buckets in configured production environments.
  • Local memory storage and mock generation are documented as development/demo paths, not as silent production fallbacks.

Evaluation and deployment

CreatorOS has focused tests around request security, schema hardening, ownership, rate-limit behaviour, timeout cancellation, mock output shape, and prompt-evaluation artifact generation. The prompt evaluator runs realistic creator fixtures and explicitly distinguishes deterministic mock output from live prompt-quality review.

Read the architecture notes ↗Read the Vercel deployment guide ↗Inspect the CI workflow ↗Browse generation and ownership tests ↗

Current maturity

This is an active prototype with a clear path toward a protected beta. The repository requires OpenAI, Supabase Auth and persistence, Turnstile, and the canonical app origin for production readiness. Without those values, the local mock path is useful for demos and contract checks but is not evidence of live model quality or hosted multi-tenant operation.

Where to look

GitHub ↗