Building LocalMate at Hack Winterthur

Source: Dev.to

Background

A few weeks back, @meritonaliu and I walked into Hack Winterthur 2026 as two CS students from ZHAW.
The City of Winterthur presented a straightforward problem: strengthen communication between the local government and its citizens.

In theory it sounds simple, but in practice people get lost in the infinite amount of documents, regulations, and complex German.

We know this challenge personally. Growing up in Zürich as children of migrants, we saw many struggle with letters they couldn’t fully understand and systems that assume prior familiarity. We often acted as translators and guides, shaping our view on information accessibility.

The Hackathon

We didn’t need to brainstorm the idea—Winterthur’s challenge mirrored the barriers we faced all our lives.

During Hack Winterthur we:

• Built a prototype in a 2‑hour train ride (11 PM – 1 AM).
• Deployed everything on a personal Proxmox server.
• Collected user feedback continuously, using document‑indexing time to let people test the tool.

The atmosphere was electric, with entrepreneurs from IT, business, fintech, and more working under the same roof.

LocalMate: A Conversational AI Assistant

LocalMate answers questions about the City of Winterthur accurately, simply, and in any language.

You can try the MVP here: LocalMate MVP

Technical Overview

• RAG pipeline (Retrieval‑Augmented Generation) with agentic workflows.
• Data scraped from stadt.winterthur.ch, ensuring information is up‑to‑date and not hallucinated.
• All documents vectorized in memory for ultra‑fast semantic search.
• Before each request, the user’s question is enriched by a small model and translated into German to guarantee retrieval quality.
• SBERT (semantic‑aware transformer) provides quick lookups that understand user intent.
• Responses are language‑agnostic: ask in Albanian, receive an answer in Albanian.

Key Features

• Verified sources only – every answer links back to an official city document or page.
• Language‑agnostic – understands semantics and reformulates answers in the user’s language, not just literal translation.
• Plain language – simplifies administrative jargon without losing accuracy.
• Agentic capabilities – beyond Q&A, LocalMate can autonomously fetch dynamic data such as schedules.

Outcomes & Next Steps

Winning Hack Winterthur was just the starting line.

• Submitted LocalMate to the WinLab co‑creating platform to collaborate with the city for a larger‑scale pilot.
• Joined the Home of Innovation incubator program to deepen business and product development skills.

The architecture is instance‑agnostic: any municipality, canton, or public/private organization with large knowledge bases could deploy its own LocalMate instance.

Team

• Anes Hodza – ZHAW student, co‑founder of Centro Labs.
• Meriton Aliu – ZHAW student, co‑founder of Centro Labs.

We publish our side projects under the name Centro Labs.

Originally featured on the ZHAW News Blog.