Skip to content

Selected work.

Systems we built and run for organizations that handle private information. We do not name our clients.

of patient records, triaged by the model
3 TB
agreement with clinicians on that triage
~93%

Triage across 3 TB of patient records.

A regional health system

The problem
About 3 TB of patient records had to be classified and triaged, and the work was being done by hand.
What we built
A supervised classifier trained on their data. It reads every record, sorts it, and puts the patients most likely to be at risk first.
Where people decide
Clinical staff work from the prioritized list instead of reading every record.
What changed
About 93% agreement with clinicians, and it cut the staff time and resources the review took.
A round control hall seen from above, rings of desks with one blue desk where a person sits
Model
Custom classifier, trained on their data
Task
Classify records and rank by risk
Review
Clinical staff, from the ranked list

Prior authorizations, prepared by an agent.

A hospital system

The problem
Prior authorizations took a large share of staff time, most of it spent gathering and checking information across systems.
What we built
An AI agent that pulls what each request needs from the health record system, internal databases, and outside sources, reasons across them, and prepares the case.
Where people decide
Staff review wherever a person needs to decide. The agent does the research; people make the call.
What changed
Much of the research and preparation now happens without staff doing it by hand.
Two white towers joined by a high sky bridge, a blue car crossing, a person looking up from below
Sources
Health record, databases, outside sources
System
AI agent with human review
Review
Staff, wherever a decision is needed

Fraud and abuse, classified by a model.

A bank

The problem
The bank needed instances of fraud and abuse picked out of its records reliably.
What we built
A language model fine-tuned on the bank's own data, which classifies each instance as fraud, abuse, or neither.
A bank hall seen from directly above, a grid of desks with one flagged in blue
Model
Fine-tuned language model
Task
Fraud and abuse classification
Data
The bank's own records

Have work like this?

Contact