Based solely on the provided article body, a self-correcting agentic loop was the key architectural change that improved reliability for Lattice, an agentic data analyst built by the UNU Campus Computing Centre. Lattice works across models and data sources behind a plain-language Q&A interface, allowing users to ask questions and receive charts, tables, or written briefs. The motivation for running models locally includes privacy, scalability, cost, and a wish to depend less on a handful of models not under the user's control. The core challenge is that the model must generate chart and query code on the fly, correctly, every time. The solution was an agentic loop that feeds a model its own errors and lets it try again. Lattice allows users to choose the model doing the reasoning and switch anytime between a local model, cost-efficient cloud models like DeepSeek or Kimi, or frontier models like Claude or GPT.