Retrieval, automation, and agent workflows aimed at a single measurable outcome.
RAG and AI agent development, out of New York. An automation earns its keep when it takes one whole job off somebody's morning. A home medical equipment provider had a dispatcher doing that job by hand: a messy export of the day's tickets, six vans, and a yard full of drivers waiting to leave.
The app runs the whole morning in four steps. It parses the delivery log and refuses to invent anything it cannot read with confidence, so an unreadable row becomes a flagged task for a human. It applies the business rules the dispatcher used to carry in their head, all of them: per-driver caps, town bans, delivery windows, driver and vehicle eligibility. Google OR-Tools then solves for minimum drive time against live traffic.
What comes out is a printed cut sheet designed for a clipboard and a pen, plus an Excel and CSV log for the office. Upload to printed routes in minutes, with every ticket accounted for. It is in daily production use on Fly.io, 1,278 tests cover the rules, and the screenshots on this site run on synthetic data because the client is not named here.
Workflow
How the routing platform runs a morningHow we work
The morning, in four steps
Parse:
The delivery log arrives as a messy export. The app reads what it can read with confidence and flags the rest for a person. Nothing gets invented to fill a gap.
Apply the rules:
Per-driver caps, town bans, delivery windows, driver and vehicle eligibility. The whole rulebook a dispatcher used to run from memory, now written down and enforced identically every day.
Solve:
Google OR-Tools solves for minimum drive time against live traffic. Geocoding fails closed, and a per-solve travel-matrix budget makes a runaway maps bill structurally impossible.
Hand over:
A printed cut sheet designed for a clipboard and a pen, plus an Excel and CSV log for the office. Six drivers colour-matched to the map, with any stop movable between them.
What it changed
Upload to printed routes in minutes, with every ticket accounted for. In daily production use, deployed on Fly.io, with 1,278 tests covering the rules. The client is not named here, and the screenshots run on synthetic data.
Who this is for
You're sitting on data or a manual process, and you want a working result you can point at. If the process is genuinely one job with a number attached to it, this is the shape the work takes.
Where this shipped
