How We Scaled an AI App Marketplace to 20,000+ Listings and 69M+ Views
Two years ago, KRAIN was a Postgres database and an idea: an AI app marketplace where discovery didn't depend on a single App Store algorithm. Today it lists more than 20,000 AI apps, has served over 69 million views, and has driven more than 21 million clicks from discovery to the tools people actually use.
None of that happened because we got lucky with a viral moment. It happened because we made a small number of unglamorous infrastructure decisions early and didn't have to re-architect around them later. That's the part of scaling a product that doesn't show up in launch tweets, and it's the part I get asked about most when founders are trying to figure out if their own AI feature is going to hold up under real traffic.
Search had to work before anything else did
A marketplace with 20,000 listings is unusable without search that understands intent, not just keywords. We built discovery on Next.js and PostgreSQL with pgvector, using OpenAI embeddings to power semantic, RAG-style search over the app catalog. That decision — Postgres plus a vector extension instead of standing up a separate vector database from day one — kept the operational surface area small enough that a lean team could actually run it. You don't need a dedicated vector database to ship good AI search. You need your existing database to do one more job well.
That single decision is the one I'd point to first for a founder asking "do we need new infrastructure to add AI search, or can we extend what we have?" Usually, you can extend what you have. The cases where you genuinely can't are rarer than the AI-tooling marketing makes them sound.
Engagement infrastructure has to survive being right
Views, clicks, reviews, votes — every one of those is a write. At marketplace scale, the write path for engagement events is where naive architectures fall over first, usually quietly, usually as a slow database rather than a dramatic outage. We built the event pipeline to handle tens of millions of engagement actions without the analytics layer becoming a bottleneck for the product layer — a separation that sounds obvious in a design doc and is genuinely easy to get wrong under a deadline, because the fastest way to ship an MVP is often to let your product queries and your analytics queries hit the same hot tables.
The lesson generalizes past marketplaces: if a feature is going to be popular, its logging and analytics need a scaling plan before its happy-path UI does, or the happy-path UI is what breaks.
Curation is an engineering problem, not just a policy one
Twenty thousand approved listings implies a rejection pipeline, spam and low-quality-submission filtering, and a review workflow that doesn't become the team's full-time job as volume grows. We treated the app-approval pipeline as a product surface with its own performance and quality requirements, not an afterthought bolted onto an admin panel. That's the difference between a marketplace that stays useful at 10x the listings and one that drowns in its own submissions.
What this means if you're scoping an AI feature
The specific numbers — 20,000+ apps, 69M+ views, 21M+ clicks — are KRAIN's. The pattern is reusable for anyone integrating AI into an existing product:
- Extend infrastructure you already run before adopting new infrastructure. Postgres with
pgvectorcovered our semantic search needs; it will cover a lot of teams' needs. - Design the write path for your busiest feature, not your most important one. Engagement logging breaks marketplaces more often than the core transaction does, because nobody budgets for it the way they budget for checkout.
- Treat moderation and quality control as an engineering surface with its own SLAs, not a queue someone checks between other tasks.
If you're not sure which of these apply to your stack, the AI Integration Readiness Checklist walks through the ten questions we ask at the start of every engagement. And if you'd rather just talk it through — we've done this work at KRAIN, and for teams from American Express to healthcare data platforms.
Metrics in this post are the sourced figures from KRAIN's internal claims register (2026-08-02): 20,000+ AI apps listed (24,216 approved in the database), 69M+ views, 21M+ marketplace clicks. Registered-user counts are deliberately omitted — they are bot-inflated by quest and airdrop farming and are not a number worth quoting.