Ride Along

What We Built

Target user: Equipment renters Customers rent CAT machinery without much background knowledge on the machine.

Pain points: the machine's history never reaches the renter. What does exist is in technical dashboards; and once they leave the lot, there's no one to ask.

Ride Along gives every machine a self-improving memory, with two sides:

For renters: choose your language (English, Spanish or French) and experience level. A "before you start" screen shows this machine's safety warnings. You can ask questions by voice and hear the answers read aloud. Answers come from this unit's history, not a generic manual that applies to everything.
For operators, technicians and fleet managers: Ride Along takes in telematics, fault codes and notes. AI infers what's happening; and stores each conclusion as a memory with its evidence. Feedback is treated as evidence, not an override, and safety warnings are never weakened by opinion.

How We Built It?

Stack: FastAPI and Postgres with pgvector, React, TypeScript, and Docker. Claude is the primary LLM, with Gemini as a fallback and a rules-only mode that needs no API key. Voice uses the browser's built-in speech tools like Google Chrome.

How it works: an evidence engine analyses a fake fleet of 10 machines with injected issues. The LLM proposes interpretations, and every claim is checked against the data. The LLM proposes, the evidence engine decides. Feedback is weighted by role (technician 1.0, manager 0.8, operator 0.6) and by each person's reliability, which is earned as later data proves them right or wrong. Answers to questions must cite stored memories, or they're discarded.

Challenges:

Keeping the LLM honest. Early on, the model would confidently cite readings that didn't exist. We fixed it using claim by claim verification on the data.
Designing feedback that matters. The memory is only as good as the last person to complain. If it's ignored, nobody trusts the system. We treat feedback as weighted evidence, which took several iterations to make.
API limits: Claude needed extra funding for an API key, so we added fallback models and a rules-only mode for other free models.

Further Implementation

Connect to real telematics feeds.
Guided pre-start checklists and a rental-counter dashboard of common renter problems.
Offline voice and more languages for job sites without reliable internet and greater access worldwide.
Adding a scanning feature for renters to check for defects using templates of the same machine.

How does Ride Along relate back to the Caterpillar prompt?

Persistent machine memory: a versioned memory of evidence, inferences and feedback for every machine.
Personalized intelligence: separate views for renters, operators, technicians and managers, tailored by language and experience.
Self-improving: new records create new memory versions, and each person's reliability is learned from outcomes.
Human + agent collaboration: people challenge the AI, it responds with evidence, and anyone can ask questions and get cited answers.
Novel interfaces: voice, read-aloud answers, and multiple languages, for the person in the seat.