Scaling at Scale: The Journey of RAIP and Genie Code at Repsol
Overview
| Experience | In Person |
|---|---|
| Track | Artificial Intelligence & Agents |
| Industry | Energy & Utilities |
| Technologies | Genie |
| Skill Level | Intermediate |
| DOWNLOAD SESSION SLIDES | |
Since launching its digital transformation program in 2018, Repsol has developed RAIP, Repsol Artificial Intelligence Products, a framework that encapsulates years of analytical expertise and supports hundreds of production-grade AI models across the organization. Within RAIP, solutions such as REPSENS for time series forecasting have been validated in demanding industrial environments.
With the rise of generative AI, rather than treating it as a separate paradigm, we focus on integrating generative capabilities with proven AI assets to maximize impact. In this session, we present ho Genie Code builds on RAIP to automate advanced workflows such as time series forecasting. We also compare this approach with foundation models such as TimeGPT and Chronos, and discuss how this agentic paradigm extends to pricing and optimization use cases. It augments existing AI models while preserving governance, reliability, and operational control and democratizes access to RAIP across business units.
Session Speakers
Victor Vaquero Soto
/Senior Data Scientist
Repsol
Full Summary
How Repsol scales predictive modeling with agents, foundation models, and internal libraries
Repsol is pushing beyond proofs of concept to make predictive modeling a repeatable, governed capability across refineries and business lines. The company combines generative AI, time series foundation models, and an agentic architecture to close the gap between subject matter experts and data scientists, in line with the unified approach described in what the Databricks Platform enables. Teams building toward this kind of scale can also find relevant context in how data and AI strategy converges on platform focus to govern and operationalize models reliably. The goal is practical: deliver faster prototyping without sacrificing production rigor. Vibe coding and zero-shot forecasting are only useful when paired with governed data, tested code, and validated models.
FAQ
RepSense is Repsol's internal Python library for time series forecasting, fully documented and deployed with CI/CD. Data scientists use its full capabilities, and subject matter experts use it through low-code interfaces. It supports use cases from fuel demand and supply chain planning to refinery soft sensors.
Every generated snippet passes through an MCP-based peer review that runs style checks, dead-code detection, and measures internal-library usage. Results are scored and tracked in a Databricks Apps dashboard so teams can monitor quality trends and address issues before deployment.
A skill packages one library capability into three files: a description of what it does and when to use it, an API reference, and real code samples. The code agent loads relevant skills before generation, which keeps outputs aligned with Repsol's best practices and library conventions.
Repsol's data is stored and governed in Unity Catalog and teams already work daily on the Databricks Platform. The platform provides the agent tooling, MLflow integration for evaluation, Databricks Apps for hosting components like the MCP server, and a consistent environment for supervised orchestration.
Foundation models like Chronos, Moirai, and TimeGPT are valuable for rapid prototyping and as strong baselines. Adoption is not uncritical due to risks such as data leakage, domain mismatch, and limited control, so Repsol focuses on combining their strengths with governed practices rather than replacing existing workflows.