Datamagnet

RevOps, sales, recruiting and market-intelligence teams need to know what a person or company looks like right now: current title, headcount, recent posts, funding. Static B2B databases go stale within months, and building live collection in-house means solving parsing, response normalisation, caching, rate limiting, per-request billing and reliability all at once. Customers also need the same response shape on every call so their pipelines never break on a missing key.

Edge and routing. A Cloudflare Worker fronts the API domain and splits traffic between two cloud backends (AWS App Runner and Google Cloud Run) with connection-level failover between them. It only retries failures that happen before the origin saw the request: credits are deducted before a response is written, so a timed-out request may already have been processed and retrying it would charge the customer twice. Every response carries a header naming the backend that served it, so incidents can be attributed.

API layer. The core is a Flask application behind gunicorn. Every data route passes through one decorator chain: authentication, rate limiting, credit resolution and deduction, then request logging and endpoint analytics. Responses are normalised to a stable shape (every documented key is always present, with null, an empty string or an empty list when a value is unknown). Callers choose a cache mode per request (serve if present, serve only if recent, or always fetch live). Accounts, API keys and request logs live in MongoDB; Redis provides the response cache and a write queue, wrapped in a circuit breaker so a Redis outage degrades to live fetches rather than errors.

Data warehouse. Profiles the API fetches are written asynchronously through a Redis queue into a PostgreSQL warehouse run by a separate FastAPI service, with batched upserts, de-duplication and a dead-letter queue. The warehouse powers the people and company database-search endpoints and the ICP filter suggestions, using derived filter columns, a materialised search index and PgBouncer in front of Postgres. An hourly aggregation job rolls raw API logs into hourly and daily analytics, and a scheduled Cloud Run job prunes the raw logs once they are folded in.

Surfaces. A Next.js customer app (API keys, request logs, billing, Signals management) authenticates with AWS Cognito; a separate Next.js admin console covers users, credit configuration, endpoint health, rate limits and provider metrics. Public docs are Mintlify with an OpenAPI spec and an interactive playground. The MCP server exposes the documented endpoints as 17 agent tools and supports both bearer-key passthrough and OAuth 2.1 with PKCE, with the customer's API key never leaving the server. Lifecycle email runs as its own always-on process so an email-provider hang can never block a payment webhook or a customer request.