In October 2025, Sandvik ran its first global hackathon: 183 participants across 42 teams, six locations, and five time zones. The brief was open-ended: build something that makes the product better using AI. Our entry was an Excel add-in that generates metrology report templates from natural language. Describe what you want to measure, and it builds the spreadsheet with the right Excel name fields, cell value patterns, and Actual/Nominal prefix system. I built the implementation side solo.
React and TypeScript inside Office.js, connected to Azure OpenAI (GPT-4o) with a RAG pipeline backed by Azure AI Search. The model had the full template spec in context, so it generated valid .XLT files that worked when imported back into the desktop application. Real-time sheet analysis, template validation, and a step-by-step workflow on Fluent UI.
// User describes what they want, AI builds the templateconst generateTemplate = async (prompt: string) => { const context = await ragSearch(prompt); // Azure AI Search const template = await gpt4o.generate(prompt, context); await excel.populateSheet(template); // Office.js await validate(template, 'verisurf-spec');};We finished as a runner-up. I conceived the project before Copilot-style agentic Excel actions started landing, but judging happened after that shift, and one judge decided the tool overlapped with something existing. The overlap was superficial: this was a metrology-specific workflow generator with the software’s naming rules, validation logic, and import constraints built in, not a general spreadsheet assistant.
The prototype later grew into VS Copilot, a full enterprise RAG system with Azure AD auth, Firestore conversations, semantic search across product manuals, and GD&T blueprint analysis via OCR, built on Next.js 15, Azure Functions, and Firebase Hosting.