WorkIPR Docket
IPR Docket
Email triage platform. About 2,500 emails a weekday, each classified within a median of 2 seconds.
106,000+
emails classified since Aug 2026
The numbers
~2,500
emails classified each weekday
2 s
median time to classify one email
0.94
AUC of the action recommender on 15,976 held-out emails
53%
of mail cleared by the action recommender
3.2%
miss rate while clearing that 53%
357
IP-law event types each email is sorted into
8,745down to58
reply templates, after clustering
16–90 sdown to1.2–2.6 s
search over 200,000+ messages
Notes
IPR Docket is an email triage platform, and I am its sole developer. It has classified more than 106,000 emails since August 2026, at about 2,500 every weekday, each within a median of 2 seconds.
Action recommender
The action recommender scores an AUC of 0.94 on 15,976 held-out emails. It clears 53% of mail with a 3.2% miss rate.
Search
Search runs over more than 200,000 messages. A query used to take between 16 and 90 seconds. It now takes between 1.2 and 2.6 seconds.
What it sorts
Each incoming email is classified into one of 357 IP-law event types and routed to the docketing team responsible for it.
How mail gets in and moves
Mail arrives through Microsoft Graph delta-sync. A PostgreSQL job queue with retry and recovery runs the work, and WebSocket updates keep the interface live as emails are classified.
Reviewers teach the model
A human-in-the-loop review flow records every reviewer decision as labelled data, so the corrections feed straight into retraining.
Reply drafting
I clustered 8,745 existing reply templates into 58 standard ones, which is what automated reply drafting works from.
Stack
Go, Vue 3, PostgreSQL and Redis. Search uses keyset pagination and trigram indexing.