07
Sweepy
llm-first inbox classification with sender cache, safe actions, and undo
Sweepy is an AI email classifier with a Chrome extension, sender cache, dual-provider fallback, action execution, and a five-minute undo system so destructive actions stay reversible.
- 2-layer pipeline
- 5-minute undo
- Claude primary + GPT fallback
Architecture
The strong part of Sweepy is the operational shape: sender cache, deterministic cost controls, provider circuit breaker, serverless-aware batch processing, and a real extension runtime around Gmail instead of only a backend classifier.
Provider strategy
Sweepy treats provider reliability as part of the product, not as an afterthought. The runtime prefers Claude Haiku 4.5 when Anthropic is configured, uses GPT-5 mini as fallback or primary fallback path, trips a circuit breaker after repeated failures, and tracks token cost so the quality-versus-price tradeoff stays explicit.
Safety model
The project is shaped around irreversible-action risk. Personal and important mail are protected, unknown stays keep by default, the prompt treats every email field as data instead of instructions, and destructive actions stay reversible through a five-minute undo window.
Serverless batching
Instead of relying on long-running backend jobs, Sweepy uses a client-driven batch loop. The scan stores Gmail message ids first, then the client processes offsets in small requests so the system can run inside Vercel Hobby time limits without queues or background workers.