James Washington
writing about archive utils free game ideas replays
Automation

Two-layer inbox classifier (rules first, LLM judge second)

Automatically labels every incoming email as belonging to one of several businesses or a personal area, so every downstream tool sees the same answer.

built Sep 2026 updated Oct 2026 Letting AI make calls, safely

how it works

Deterministic recipient-domain rules from a single taxonomy JSON decide first; an LLM classifier only handles the remainder, low confidence goes to a Triage/Unsure label, and every verdict is written to an evidence ledger.

Automation: Something that runs on a schedule by itself.

The code for this one lives in a private repo. This page explains the idea so you can build your own.

classificationemaillaunchdllm-judgerules-first

related

  • Code-owns-workflow, models-supply-judgment outreach pipeline Automation A program that researches leads and writes draft sales emails for a person to approve, never sending anything itself.
  • Single-source org/sphere taxonomy read by every agent Document One file that defines which business or life area something belongs to, used by the email sorter, the task board and the agents alike.
  • Template-only auto-replies as a prompt-injection defense Technique An auto-responder that never lets an AI read the incoming email can't be tricked by it.
  • agent-merge: Jev-gated self-merge for agent branches Automation The one sanctioned way an AI may merge its own work: a script that reviews it first and refuses if it didn't show its work.
  • Run a classifier in shadow mode before it can gate Technique Let a new AI judge watch silently until there's evidence it's right.
← everything I've built
© 2026 James Washington
github twitter rss