# *AURORA* A complete funding proposal for an AI-supported emergency alert system: from broadcast warnings that tell everyone the same thing to location-precise, explained and actionable messages. Planned as a four-partner consortium over 24 months with a €1.97M budget. > Project page: https://clemenshoeller.at/en/projects/aurora > Category: consulting/funding-proposal ## Key facts - **Format** · Consortium proposal - **Course** · Business Project - **Institution** · WU Vienna, master's - **Period** · Summer 2026 - **Work mode** · Group of four - Proposal as PDF ↗: https://clemenshoeller.at/projects/aurora-project-proposal.pdf ## /01 The question Austria's AT-Alert system reaches very many people very quickly, but everyone gets the same message. Someone standing in the middle of the danger zone reads the same text as someone fifty kilometers away. At the same time the events are becoming more dynamic: more extreme-rainfall days, more heat days, the 2024 flood alone causing around €1.3 billion in damage. A warning that sounds the same for everyone creates panic in some and habituation in others. Research calls this effect alert fatigue: people who are warned irrelevantly too often stop paying attention. The goal was not to build this system but to do what comes before that in the real research world: a complete, submission-ready project proposal of the kind that programs like Horizon Europe or the Austrian FFG require. Concept, state of the art, methodology, work plan, budget, consortium, risk management and exploitation, all in one document that would have to convince a funding body. Our answer is **AURORA** (AI-supported Unified Risk Observation and Response Assistant): a system that does not replace the existing warning infrastructure but sharpens it. ## /02 The concept The ideation started with an observation about the state of research: all the building blocks for better warnings already exist, nobody has just connected them. Alert systems broadcast fast but do not know whom they reach. Geospatial risk models know the hazard zones but do not talk to the people affected. Language models phrase things well but hallucinate without an anchor in data. AURORA's core idea is therefore not a new building block but the missing connection between them. That turned into an architecture of three layers: - **Knowledge graph:** An RDF-based graph integrates weather data (GeoSphere Austria), geospatial and infrastructure data (OpenStreetMap) and real-time events, queryable via SPARQL. The system then knows not just that there is a flood, but which places, roads and facilities depend on it. - **Risk layer:** Rule-based geospatial analysis (buffers, overlays, thresholds) classifies affected people into low, medium or high risk. Deliberately not a black box: every classification can be traced back to explicit criteria, which the EU AI Act requires for safety-critical AI anyway. - **Grounded LLM:** A language model translates the structured risk information into clear, actionable warnings. It only phrases what the graph and the risk layer substantiate, following the retrieval-augmented generation principle that curbs hallucinations. The most interesting design decisions lie in what AURORA deliberately does not do. No machine learning model in the risk layer, tempting as it would be: in a disaster, a black box cannot explain to anyone why it warned them or failed to. No delivery channel of its own, but a layer on top of the mobile network's cell broadcast: the path to people's devices stays the one that already works today, only who gets which message becomes more precise. And no freely generating language model, but one kept on the short leash of the data. ### Three measurable objectives Funding proposals live on verifiable commitments, not visions. The proposal therefore defines three SMART objectives: by month 7, a queryable knowledge graph built from at least three data sources, measured against a catalog of 20 competency questions with at least 95 percent answered correctly. By month 10, a risk assessment that updates risk zones within two minutes of new hazard data, evaluated against expert-annotated ground truth. And by month 18, an integrated prototype competing against a reconstructed AT-Alert baseline in at least three realistic scenarios (flood, severe thunderstorm, wildfire), measured by precision and recall of affected-user identification and the share of irrelevant alerts. Methodologically the proposal follows design science research: design, build, evaluate, refine, with stakeholder feedback in every cycle. ## /03 The proposal The actual deliverable is the management scaffolding around the idea: five work packages totaling 258 person-months, a bottom-up budget of €1.97 million, a four-partner consortium and procedures for risk, data and legal compliance. - WP1 Management & dissemination *runs the full duration*: PM - WP2 Knowledge graph *data foundation, starts first*: PM - WP3 Risk assessment *developed in parallel with WP4*: PM - WP4 Grounded LLM *communication layer*: PM - WP5 Integration & validation *scenarios against the baseline*: PM *Person-months per work package, 258 in total over 24 months (2027 to 2028). The most expensive package is deliberately the validation: a prototype that never competes against the baseline would just be a claim.* ### Budget: bottom-up, not a guess The budget is built bottom-up from roles, person-months and researched salary levels (FWF rates for research staff, market salaries plus 20 percent employer costs for the rest). On top of the planned activities of €1,749,247 comes a contingency of €221,531, for a total of €1,970,778. - Personnel*258 person-months, 4 partners*: 73% - Infrastructure*compute, hosting, geo APIs, testing*: 12% - Contingency*project-level reserve*: 11% - Other*dissemination, open access, workshops*: 3% - Travel*plenary meetings, EU events, validation*: 2% *Total budget of €1.97M by cost type. The contingency of roughly 12.7 percent of the planned budget is fairly moderate for an integration project with four partners.* ### Consortium and governance Four partners, four necessary capabilities: TU Wien as scientific lead (knowledge graphs, semantics, evaluation), Utimaco for secure implementation (already builds alerting systems in Austria), A1 carrying the delivery via cell broadcast in its own mobile network plus project coordination, and the Ministry of the Interior as the end user bringing civil-protection requirements, approvals and validation. Add the procedures a proposal of this size needs: PRINCE2-based risk management with a central register, IP governed by a DESCA consortium agreement with a selective open-source strategy, data management along the FAIR principles and the legal mapping across GDPR, the EU AI Act and the EECC. The system stores no personal data; location information is processed in aggregated form only. ## /04 Context The proposal took shape within the Digital Economy master's program, in a format that frames project management for research-oriented ventures explicitly along the logic of Horizon Europe and the FFG. Topic and idea were our own: we pitched AURORA, sharpened it in several feedback rounds with domain experts, and built it out over a semester, the four of us, from the first pitch to the full proposal. > **Course project** — AURORA is coursework, not a submitted proposal: the consortium is hypothetical and the organizations named were not involved. The ambition was to write it as if it had to convince a real jury nonetheless, with real data sources, real salary rates and the real legal situation. ### What I take from it The system idea was the smaller part of the work. The larger part was making it fundable: distilling three verifiable objectives from an architecture, work packages from objectives, person-months from work packages and, from person-months, a budget that survives scrutiny. That translation chain, from idea to defensible plan, is what I take from the course. **Skills gained:** Project planning (TCQ) · Budgeting · Risk management (PRINCE2) · Funding logic (Horizon, FFG) · System architecture · Academic writing ## /05 Related projects These projects connect more tightly than the tiles on the home page suggest. Here are all the others, sorted by how closely they relate to this one: - [Prompt Injection](https://clemenshoeller.at/projekte/prompt-injection) - [Healthcare Fraud Detection](https://clemenshoeller.at/projekte/fraud-detection) - [AI Talent Gap](https://clemenshoeller.at/projekte/ai-talent-gap) - [Consulting Cases](https://clemenshoeller.at/projekte/consulting-cases) - [Zotter Marketing Plan](https://clemenshoeller.at/projekte/zotter-marketingplan) - [AI Automations](https://clemenshoeller.at/projekte/ki-automatisierungen) - [LLMs & Creativity](https://clemenshoeller.at/projekte/llms-kreativitaet) - [Google vs. ChatGPT](https://clemenshoeller.at/projekte/google-vs-chatgpt) - [This Website](https://clemenshoeller.at/projekte/diese-website) - [VWA: Origin of the Internet](https://clemenshoeller.at/projekte/vwa) --- *German version: https://clemenshoeller.at/projekte/aurora.md* *AURORA: an AI-supported system for location-precise, explainable emergency alerts, developed into a complete funding proposal (WU Vienna, master's). Knowledge graph, risk layer and grounded LLM, planned as a consortium with a €1.97M budget.*