Personal tool · 2026 · Practical AI
Junkmail AI Cleaner
A public tool for a repetitive mail workflow: identify known spam quickly, use a local model for uncertain messages, and keep the process recoverable when classification or dependencies fail.
The problem
Reviewing an Outlook junk folder takes repeated attention. Automated deletion raises a harder question: how to reduce junk without silently losing legitimate mail.
Song’s contribution
I published the tool and its implementation. The repository combines keyword rules, Microsoft Graph access, local Ollama classification, scheduled runs, and a preflight check that stops a run if required services are unavailable.
Approach and evidence
Known patterns take the fast path. Uncertain messages are classified locally. The documented output contract treats malformed model responses as “keep”, and Graph deletion moves mail to Deleted Items so a mistake can be recovered.
The public repository includes unit tests, model-output tests, and a labelled evaluation set. Its README states a zero-deleted-legitimate-mail gate and documents known gaps; this page does not claim a measured production result.