dérive

Consumer layer

Users discover, save, and revisit curated food & drink spots. A taste profile is built from onboarding preferences and behavioral signals — saves, return visits, browse patterns. Subscription model via Stripe, RevenueCat, and Apple Pay with full entitlement lifecycle management.

Vendor portal

Live vendor-facing platform at vendor.ourderive.com where restaurants, bars, and venues claim and manage their dérive profiles. Vendors upload menus and brand lists — PDFs, images, or documents — which are processed by Claude Vision and stored at item level in the dérive database.

Agentic trip planner

Users describe what they want in natural language. Claude extracts location and intent, queries item-level menu data, and returns a personalized itinerary with contextual notes explaining why each spot matches this specific user's taste profile.

Taste agent

Claude-powered onboarding that infers structured preference tags from free-text input — cuisine preferences, drink style, mood, pace. Tags feed directly into the recommendation engine and trip planner from the first session.

Vision — menu parsing

The core data pipeline. Claude reads vendor-uploaded menus — any format, any layout — and extracts structured item-level data. No templates, no rigid parsing rules. Handles handwritten specials, wine lists, tasting menus, and brand sheets with the same pipeline. This was the technical bet that made the vendor portal viable.

Edge functions — production agents

Two Claude-powered agents shipped in production: the Taste Agent infers preference tags from free-text user input; the Trip Planner queries vendor item data combined with the user's taste profile to generate personalized itineraries, with programmatic pre-filtering to ensure must-match spots always surface.

Experiment design

Used Claude to generate onboarding hypothesis variants, pressure-test measurement plans, and draft user research scripts before writing a line of code. The A/B test design, success metrics, and guardrail definitions were all iterated in conversation first.

Prompt engineering

Iteratively designed and tested prompts for each agent — defining output format, edge case handling, and the tone that makes recommendations feel curated rather than algorithmic. The difference between a good trip planner and a great one lives in the prompt.

Build acceleration

SQL for behavioral cohort analysis, React component boilerplate, Supabase schema migrations, API route handlers — moving from spec to working code significantly faster than the team size would otherwise allow. This entire platform was designed and built solo.

Vision parsing is production-ready

Vision parsing is production-ready. Claude Vision handling menus at item level — across formats, layouts, and handwriting — without structured templates was the technical bet that made the vendor portal viable. It worked.

The data layer is the product

The data layer is the product. The trip planner agent is only as good as the item-level data underneath it. Investing in the parsing pipeline first was the right call. Great models can't fix thin data.

Agentic systems need tight tool design

Agentic systems need tight tool design. The quality of the trip planner response is a function of how precisely the database query is structured before Claude gets the results. Garbage in, garbage out — even with a great model.

The PM takeaway

The PM takeaway. AI compresses the time between idea and testable artifact — but the judgment calls about what to build and why stay human. Claude is the accelerant; product thinking is still the engine.