Reads everything. Remembers nothing: an email watch API built on open-weights AI

WatchFox is a new email-watching API that allows users to connect their mailboxes and set up rules for specific types of emails such as enquiries, invoices, support tickets, purchase orders, job applications, claims, or custom types. When a matching email arrives, WatchFox instantly sends the information to a server via a signed webhook, without having to read the inbox or train a classifier.

The SaaS builders and operations teams that can benefit from this technology include sales teams, accounting departments, support teams, and procurement teams.

WatchFox solves the problem of teams having to write and maintain complex email parsing and classification systems, which often involve IMAP polling, regex rules, and LLM prompts. By providing a single API call, WatchFox simplifies the process and addresses privacy concerns by never storing email content. It achieves this by using an open-weights decision model called Laya typed-decisions, a 421M-parameter encoder that answers typed, calibrated questions about a piece of text with yes/no probabilities, choices, and ordinal scores.

The WatchFox API is built using FastAPI, React, PostgreSQL, Azure App Service containers, and GitHub Actions. The codebase consists of around 6,400 lines of Python, 3,500 lines of React, and 74 automated tests that run on both SQLite and PostgreSQL. The service has been tested under load, with 100 accounts and 5,000 emails processed in approximately 31 seconds while handling 50 concurrent API requests with a median response time of 0.3 seconds.