Frankie — Credit Card Intelligence Platform

Distilled PRD · v0.2 · 2026-08-30. The full original brief lives in the founder's notes; this document is the working source of truth.

One-liner

The intelligence layer for the credit card industry: a structured, continuously updated, historically versioned database of US consumer credit card products — fees, APRs, rewards, bonuses, benefits, transfer partners — with provenance on every field.

Thesis

Credit cards are unusually complicated products (dozens of attributes, changing constantly), yet the information lives in marketing pages, PDFs, and affiliate listicles. Nobody treats it as structured, historical, queryable data. Whoever does becomes the source of truth — see research/competitors.md: the "Bloomberg for credit cards" lane is unoccupied, and history is the moat a late entrant cannot backfill.

Model company: Channel 3 (trychannel3.com) — "the world's highest quality product graph, so AI agents can reliably search, recommend, and purchase products," one API over 100M+ products, sold to developers, brands, and enterprises. Frankie is that playbook for credit cards: the canonical card product graph that analysts, fintech apps, and AI agents query — with the twist that our graph is also historical and provenance-carrying, which commerce catalogs don't need but financial intelligence does. As agentic commerce reaches financial products ("which card should I get/use?"), the AI answering that question needs a structured, verified card database — Frankie is that layer.

Two applications, one asset

• B2B market intelligence (first): search, compare, benchmark, and track changes for issuers, fintechs, loyalty teams, consultants, investors.
• B2C AI advisor (later): recommend optimal card portfolios from a spending profile, powered by the same engine.

Positioning rule: we are not "AI that recommends credit cards" — that's one application. The company is the data + intelligence layer.

Product architecture (four layers)

1. Data — normalized card database (see schema below).
2. Intelligence — competitive sets, historical changes, benchmarks, trends, market events feed.
3. Modeling — simulate product changes ("$495 fee + 4× dining vs today?"). Post-MVP.
4. Applications — B2B dashboard, B2C advisor, API.

MVP scope (what this repo builds)

Seven screens, mock data, no backend:

The MVP question: will professional users pay for a dramatically better way to search, compare, and analyze credit card products?

MVP non-goals

Perfect coverage, bank integrations, transaction ingestion, ML prediction, consumer mobile apps, automated applications, international coverage, authentication (added with real data), and the simulation engine.

Data model (domain types in packages/shared-types)

Issuer · CreditCard (fees, AprProfile as ranges — never single numbers, SignupOffer, RewardRate[] at category level with caps/portal flags, CardBenefit[], transfer partners) · CardEvent (typed change records with old/new values and a competitive direction — changes are stored, never overwritten) · Provenance (source type, URL, verified date, confidence) on every card and event.

Full relational schema for the Supabase phase: issuers, networks, cards, card_terms (versioned), signup_offers, reward_categories, card_rewards, benefits, card_benefits, transfer_partners, card_transfer_partners, sources, data_evidence, card_events.

Product principles

1. Structured first — facts in fields, not prose.
2. Historical by default — updates create events, never destroy state.
3. Source everything — every number traceable (source, date, confidence).
4. Separate facts from estimates — verified / estimated / modeled are visually distinct (mock data is labeled as such in the UI).
5. Human-readable + machine-readable — same data powers UI, exports, API, AI.
6. Intelligence over aggregation — the value is what the data means.

Roadmap phases

1. Data foundation — reliable structured data for ~100–250 major US cards → plans/card-data-ingestion.md.
2. Intelligence dashboard — this MVP + auth + exports + Supabase.
3. Change tracking — automated refresh + market changes feed from real diffs.
4. AI analyst — natural-language queries compiled to structured filters; answers from the database, never model memory.
5. Consumer MVP — questionnaire → expected-annual-value recommendations (formula already implemented in lib/domain/valuation.ts).
6. Simulation — enterprise scenario modeling.

Monetization (hypotheses, validate via discovery)

B2B SaaS: analyst ($500–1k/mo) → team ($2–5k/mo) → enterprise ($25–100k+/yr with API, exports, historical data, alerts, simulation). Consumer: affiliate (rankings never influenced by payout — disclosed), premium subscription. API as its own product line.

Success metrics

Data: cards tracked, fields/card, % primary-source provenance, freshness. B2B: active orgs, weekly active analysts, searches, comparisons, retention, ARR. Consumer (later): recommendations completed, conversion, MAU.

Top risks