CampFit: Denver Kids Camp Discovery

The Problem

Every Denver parent planning a summer ends up with the same thing: a sprawling spreadsheet of kids’ camps, half of it stale, none of it comparable. Registration windows open and close before you have finished reading. I had that spreadsheet too, and I wanted to turn it into something other parents could actually use. The hard part was never the UI. It was trusting the data.

What I Built

CampFit is a Next.js, Prisma, and Supabase progressive web app where parents can discover, compare, save, and get notified about kids’ camps. The part I care about is underneath it.

• AI-driven crawl and extraction: camp information lives in hundreds of inconsistent websites, PDFs, and registration portals. I built an agentic pipeline that crawls those sources and extracts structured fields (sessions, ages, pricing, locations, registration dates) into a normalized model, instead of me maintaining a spreadsheet by hand.
• A trust and review pipeline (the point of the whole thing): nothing a crawler finds is shown to a parent until it earns it. Data flows through crawl, then extraction, then a proposed field claim, then human review, then an approved value with a public trust signal explaining why it should be believed. The review workbench is where AI-gathered data has to prove itself before another parent relies on it.
• Domain-driven core: the codebase is organized around bounded contexts (Camp Discovery, Data Stewardship, and Trust and Review Provenance) with a documented ubiquitous language, so the model stays honest as it grows.
• A reusable provenance layer: the review pattern is not camp-specific. It runs on a general provenance layer, which is proof that one trust-and-review approach generalizes across very different kinds of data.
• Real delivery discipline: Playwright production smoke tests in CI, architecture decision records, and a freemium model (premium alerts, unlimited saves, calendar export) sketched out on Stripe.

Why It Matters

CampFit is the human-scale counterweight to my infrastructure work, a real product for real parents in my own city, but it is built on the same conviction I bring to enterprise AI: you do not trust agents, you build systems that make them prove their facts. The fun story is that AI helped assemble a working app in a weekend. The story I actually care about is the crawl-extract-review pipeline that decides which of that AI-gathered data is trustworthy enough to put in front of another parent.

Takeaways

• A directory is only as good as the trust behind each fact in it, so the review layer is the product, not a nice-to-have.
• The same provenance pattern that vets crawled camp data also vets very different sources, so the approach is domain-agnostic.
• Building for your own community keeps the feedback loop honest.