Work
Recent · named clients
4h 30m → 35mPer day, rekeying
Halden FreightFreight brokerage · 40 people
Two dispatchers spent most of a morning copying rate confirmations out of emailed PDFs and into the dispatch system. We built something that reads the PDF, finds the matching load, writes it in, and holds anything that does not line up for a human. It has been running unattended since March.
Six weeksPython, Postgres, their existing TMS API
They own itTheir repo, their AWS account
3 days → same dayPrior authorisation packets
Corven HealthTwelve-clinic group
Assembling a prior authorisation packet meant a coordinator reading a chart and hunting for the four documents an insurer wanted. Now the packet assembles itself and flags what is missing before it goes out. First-pass denials dropped by a third, which was the number the practice manager actually cared about.
Eight weeksRuns inside their EHR vendor's API
ReviewedEvery packet still gets a human signature
2 days → 3 hoursQuote turnaround
Ravensworth BrokersCommercial insurance
Brokers were comparing policy wordings from six carriers by opening six PDFs side by side. We built a comparison that pulls the clauses that differ and shows them next to each other, with a link back to the page it came from so nobody has to trust it blindly.
Five weeksSearch across their own document library
Nothing inventedEvery line cites its source page
9h → 40mFirst response time
TallowDirect-to-consumer, 200k customers
Support was three people and a backlog. The system now drafts a reply with the order, the shipping status and the returns policy already pulled in. A person reads it and sends it. Seventy-one percent go out with no edit at all, and the three people still answer the hard ones.
Four weeksSits inside the helpdesk they already paid for
Never auto-sendsA person presses send, every time