AI Workstreams
Project Update
April 23, 2026
Confidential
Apr 10 · What's Next → Status
Four items. Here's where each landed.
Alias custom app URL to box.fpamed.com
Not done
Need access to fpamed domain registrar to create the record. 3-minute change once access is in hand.
Run live workflows with real cases
Done
8 depo summaries + 1 index run on Apr 20, self-serve by Jacqui. Details on the next slides.
Begin Med Chron validation testing
In progress
Dev environment setup underway. Spike: validate Box AI against real source material before building pipeline.
Settle on Box Consulting's role
Done
JH: ad-hoc custom dev on med chron. FPA: all other Box consulting needs.
Live Production Run · Apr 20
Both workflows. Self-serve. No assistance.
Depo Summary — 8 Jobs
Avg pages / transcript54
Avg runtime1.43 min
Processing speed1.67 sec / page
Document Index — 1 Job
Files processed792
Total runtime49.6 min
Processing speed3.76 sec / file
Depo Summary · User Feedback (Mark)
Strong signal on quality.
One real fix needed.
"At first glance, the Depo summary is amazing."
"As far as I'm concerned, the contents of the Depo summaries are absolutely fine."
"My time is gonna be much reduced."
Issue
Transcripts submitted were condensed format: 4 physical pages per page. PDF links in the summary navigate to the condensed page, not the actual testimony page within it. Mark flagged this as substantive, not cosmetic.
In progress: building accurate link navigation for 4-up condensed transcripts.
Document Index · Efficiency
792 files. 50 minutes.
Manual would have been 6-7 hours.
~90%
reduction in per-file processing time
3.76 sec / file automated vs. 30-45 sec / file manually
~98%
reduction in total labor time
50 min elapsed (automated) vs. ~6-7 hours of staff time
Staff time freed from catalog work is reallocated to higher-value activity. Where labor cost is billed to clients, the delta is margin.
What's Next
One fix in flight.
Then all focus on med chron.
In flight
4-up condensed depo link accuracy
Building navigation logic so summary links target the correct testimony page within a condensed (4-up) transcript PDF, not just the condensed page number.
Up next
Medical chronology validation sprint
Confirm what the source material looks like and what the output needs to be before writing a line of pipeline code. Box AI against real medical records.