Transit Routing with Jev and Neo4j: A Typed Decision Model in the Graph

Session track: AI Engineering

Session time: 07:00 am - 07:30 am UTC November 12

Session description:

Language models are good at writing text, but many applications don't need text. They need a decision: take this branch or that one, approve or escalate, route here or there. Jev, a System One model, answers a question with a typed choice and a probability for every option, so you can see how sure the model is and not only what it said. City transit is a natural fit. When stations close or delays build, a rider or an operations team has to choose between routes and a wrong call is costly. We use one such decision from the London Underground Northern line, which branch to take when stations are suspended or delayed, as a worked example of a pattern that applies to any transit network and show it two ways. First from Python, in a few lines of code. Then inside Neo4j, where the state comes from the graph, the decision is a single Cypher call and the same graph finds the way around. You’ll see the same call run against a hosted model and a small local one, and what to do when the network is gone. Key takeaways - A decision that comes with probabilities and a margin is easier to trust, test and escalate than free text. - The same pattern works for any transit network that can be described as a graph. - A hosted model and a small local model can answer the same question through the same call. - Putting the decision next to the data, in the graph, removes a layer of glue code.