Built a production AI SaaS that researches products across Reddit, TikTok, Google, and YouTube, then generates buyer personas, platform-specific ad copy, and ad imagery, with Stripe billing and full observability.
Reddit, Google, TikTok, Facebook, Instagram, and YouTube
Bounded headless Chromium pool, throughput without memory collapse
Stripe subscriptions alongside one-time credit purchases
Ad creative starts with research nobody wants to do: reading what actual buyers say across half a dozen platforms, turning that into personas, then writing platform-specific copy against each one. Doing it properly takes days; skipping it produces the generic creative that fails.
Built a SaaS where a user pastes product URLs and the system runs a multi-stage research and generation pipeline, scraping across platforms, then chaining Claude models to produce personas, copy, and images. No ad account connections: users download the output and take it to their own ad manager.
The generation pipeline is a chain, and not every link needs the same model. The cheap, high-volume first chain runs on Claude Haiku; the chains that produce personas and final copy run on Sonnet. Cost tracks the work actually being done rather than the most expensive step in the chain.
Scraping and generation are minutes-long, failure-prone operations, so they run as BullMQ jobs against Upstash Redis with a bounded three-worker Playwright pool. The web request returns immediately and the browser subscribes to progress rather than holding a connection open.
Job progress is published to Redis and fanned out to the browser over Server-Sent Events, so a multi-minute generation feels observable instead of frozen, and a reconnecting client resumes from the current state.
Image generation sits behind an ImageService interface with OpenAI GPT Image and fal.ai Flux implementations. Model quality and pricing in this space move monthly, the abstraction means switching providers is a config change, not a refactor.
Workers publish progress to Redis channels; the Next.js route handler subscribes and forwards to the client as SSE. Any worker on any container can drive any user's progress stream without the two ever knowing about each other.
redis.subscribe(`job:${jobId}`, (msg) =>
controller.enqueue(`data: ${msg}\n\n`)
);Sentry for errors and PostHog for product analytics were wired in before launch rather than after the first incident, in a pipeline this asynchronous, a failure with no trace is a failure you will never reproduce.
If something here looks like what you're facing, tell me about it. The first conversation is an architecture discussion, not a sales call.
