JH
Jonathan Haber

What has been built / What you'd have

Contents

What you'd have

Systems that learn what actually works

Quality measured from what people do rather than what a model says about itself, and engines whose output is checked against real results.

Quality measured from real behaviour

Coaching quality tracked from what members actually do (how long they stay, how deeply they engage, whether they return), with nightly alerts when it regresses.

What it gives you

An early warning that reflects real experience, not a model grading its own work.

What makes this version different

  • Built after a period when an automated quality score said all was well while real engagement was falling for weeks. Monitoring was moved onto behaviour as a result.
  • A separate, rubric-based evaluation is kept for research, not trusted as ground truth.
Approx. lines
~4,900
Since
Mar 2026

Measured 26 September 2026

An outreach engine with a human gate

Personalized lifecycle emails written by AI (check-ins, re-engagement, digests), with a human approval queue before anything is sent and rules that respect bounces, unsubscribes and spam reports on every send.

What it gives you

Outreach at the scale of AI with a person still deciding what goes out.

What makes this version different

  • A per-member send record designed to keep cadence and prevent duplicates, and a dedicated test suite for its safety rules.
  • Explicit, written rules for when human review is required.
Approval-system commits
183
Approx. lines
~21,000
Emails logged
25,638

Measured 26 September 2026

A search and answer-engine content engine

A pipeline that finds what people search for, screens topics for safety, generates and maintains more than 1,300 pages, checks their health every day and reports in plain language.

What it gives you

An organic channel steered by search data you own.

What makes this version different

  • Screens out medical and crisis topics before anything is written.
  • A daily scheduled health check that catches its own regressions.
Automation scripts
383
Commits
356
Pages generated
1,380

Measured 26 September 2026