Luis Paredes Franco
← Projects

Five AI products in six months at VEC Fleet

AI Product Owner · VEC Fleet · Mar 2026 – Present

I co-define the AI roadmap for a fleet management platform. Three AI products shipped and two more are in development, including a conversational analyst that answers operational questions from client data.

3 + 2
AI products shipped + in development
20
clients live on the data assistant
83%
answer acceptance rate

1Context

VEC Fleet is a fleet management SaaS platform. I joined to own the AI initiative backlog: finding the opportunities, defining each product, and getting it shipped with the engineering team. The roadmap is co-defined with the CPO. I propose initiatives and the success metrics for each one, and we agree priorities together.

I had used the platform before I worked on it. At AB InBev I was the customer who implemented it. That is why most of these products start from an operational problem I have seen first-hand.

2The portfolio

ERIC Analyst (shipped). A conversational assistant inside the platform. It answers questions about how to use the product, and it connects to a client's database through SQL to answer questions about their own operation. Clients can also upload their documents for retrieval-based answers.

Predictive engine-damage model (shipped). Telemetry and several separate databases unified into one source, with a predictive alert layer on top. Rolled out to a large client in six stages, starting from a single pilot site.

ERIC Inspector (shipped). A computer vision model that detects vehicle faults from video.

ERIC Compliance (in development). An OCR pipeline that reads driver documents and credentials so they can be bulk-loaded instead of typed in by hand.

ERIC Discoveries (in development). Irregularity detection. A deterministic rule set flags anomalies, then an LLM writes the client-facing report that explains what was found and what to do next.

3Decisions I would defend

Two capabilities, two rollouts. Answering "how do I use this screen" and answering "which of my vehicles cost the most last month" carry very different risk. The first went live for every client. The second touches real client data through generated SQL, so it went to 20 clients first.

Two ways of measuring quality. The 83% acceptance rate for ERIC Analyst comes from user thumbs-up feedback and from human expert review of answers. Either one alone is easy to fool. When they agree, I trust the number.

One threshold per vehicle model. The predictive model does not use a single global decision threshold. Failure signatures differ by engine type, so each vehicle model has its own tuned threshold. A missed failure is expensive and a false alarm costs an inspection, and that trade-off is not the same for every truck.

Rules decide, the LLM explains. In ERIC Discoveries the detection is deterministic and auditable. The language model only writes the explanation. A client can always ask why something was flagged and get the same answer twice.

4How I work with the team

I run sprint planning and refinement with the developers in Jira, speak with clients directly, and prototype improvements myself with Claude Code alongside the technical lead. For ERIC Analyst I also led the competitive research and proposed the adoption strategy.

5Enablement

I designed and delivered an eight-session AI training programme for about 40 people across Sales and Marketing, Product and QA. It covered AI fundamentals, Claude, Claude Code, MCP and agents. The sessions turned into four working tools: QA agents that run test sessions, a research agent for sales prospecting, proof-of-concept skills for the product team, and an agent that drafts the initiative backlog from client orders.