James Washington
writing about archive utils free game ideas replays
Automation

Code-owns-workflow, models-supply-judgment outreach pipeline

A program that researches leads and writes draft sales emails for a person to approve, never sending anything itself.

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

how it works

Stdlib Python where code does fetching/rule checks (word count, one ask, no hype), an LLM classifier gates the hook, a cheap model drafts, a model picks a contact by INDEX into a code-built list so it cannot invent an address, and a second model reviews; output is a Gmail draft plus a single approval task. Includes threshold calibration against hand-labelled leads and a model bench.

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.

anti-hallucinationcalibrationhuman-in-the-loopindex-not-valuellm-pipeline

related

  • Two-layer inbox classifier (rules first, LLM judge second) Automation Automatically labels every incoming email as belonging to one of several businesses or a personal area, so every downstream tool sees the same answer.
  • Attorney-gated legal document pipeline App Drafts legal paperwork with AI but routes every document through a licensed attorney before anyone receives it.
  • Marked-up LLM gateway with per-client virtual keys and Stripe invoicing API A private AI service that lets a consultancy give each client their own AI key and bill them monthly for exactly what they used.
  • AI Workbench: local-first LLM interface Browser tool A chat workbench in one web page that talks to local or cloud AI models.
  • Fleet run grader + Langfuse calibration Automation Every night, grades the unattended AI workers' runs against a rubric, and asks James to spot-check a sample so the grader stays honest.
← everything I've built
© 2026 James Washington
github twitter rss