SEMINAR · UNIVERSITY OF ZURICH
Applied AI Systems: Architecture, Integration, and Scaling
Applied AI Systems is a bachelor’s seminar (3 ECTS, in English) that Dr. Oliver Gausmann has led at the University of Zurich’s Department of Informatics since fall 2026. Over twelve weeks, small teams build a working proof of concept for a real case and back their architecture decisions with cost, risk and value.
What it is about
The seminar covers how an AI prototype becomes a working system that’s built with cost in mind, and how it fits into an organization and reaches a market.
Four of the cases come from industry partners who work with the teams directly: they present the case at the kick-off, answer questions in an online session during the build phase and join the final presentation. The other three are the seminar’s own cases.
Its structure follows the focus of Oliver Gausmann’s own work with software companies and investors: making engine-room decisions on models, data pipelines and operations that hold up in front of a board. So the teams work through the questions a board would ask: what the system costs, and who’d buy it.
How the seminar works
Frame and design
Frame the problem on one page, sketch the architecture and record the first decision as an architecture decision record. Teams bring all three to the design review.
Build and harden
A running end-to-end slice with data, model and a usable result, plus a risk and cost log. Both go into the mid-point demo.
Evaluate and present
Evaluation results with a documented method, a written report that sets out the architecture rationale and a view on go-to-market and adoption, and a final presentation with a live demo.
Topic pool
The fall 2026 topic pool has seven cases: four from industry partners and three of the seminar’s own.
Industry partner case
Brand Health Engine
Measuring from the outside how much trust a brand earns, with evidence per finding, a score and a trend over time.
Methods: Multi-factor reputation measurement · LLM as a rater, calibrated against human ratings · AI answer engines as a source
Industry partner case
House-Style Generation on Open-Weight Models
A self-hosted open-weight model learns a house style and is benchmarked against API models.
Methods: Retrieval versus fine-tuning with LoRA · Evaluation with blind ratings · Unit economics and data sovereignty when self-hosting
Seminar case
AI Cost Control Layer
A control layer between application and model provider meters every AI call and checks it against budget, cost target per task and thresholds.
Methods: Cost per solved task as the metric shared by engineering and finance · Gateway and policy pattern · Twelve-month forecasts and budget limits
Seminar case
AI Learning Trainer for School Tests
From a one-page concept to a verdict: is a learning trainer that builds its exercises from photos of a pupil’s own workbooks technically feasible and commercially viable?
Methods: Structured extraction with vision models · Confidence values and human review · Minimal-data design for a product used by children
Industry partner case
Consumer and Market Signals for Product Innovation
Continuously collect consumer, society and market signals and turn them into innovation opportunities for product management.
Methods: Separating weak signals from hype · Semantic clustering and momentum analysis · Explainable multi-criteria prioritization
Seminar case
Scalability Scan for B2B Software
An agent reads a software company from the outside and scores how scalable it is and how it looks to a buyer, with evidence for every statement.
Methods: Research agents with stopping rules and cost per scan · Citation checks against hallucination · Limits of outside-in analysis
Industry partner case
Internal Data Analysis on Self-Hosted Open-Weight Models
A self-hosted open-weight platform lets a company analyze its internal data with a language model without the data leaving the company.
Methods: Self-hosted open-weight model · Retrieval, with fine-tuning only where needed
What companies can take from it
- Every team keeps a cost and risk log and justifies its architecture on scalability, cost, governance and risk. Depending on the case, a metric of its own comes on top, such as cost per solved task, per 1,000 tasks or per scan.
- Whether a self-hosted open-weight model or an API model is the better fit comes down to numbers: quality, response time and cost per 1,000 tasks, plus the question of which data may leave the company.
Key facts
- Format
- Seminar, bachelor’s level (BSc)
- Credits
- 3 ECTS, twelve weeks of project work
- Language
- English
- Department
- Department of Informatics, University of Zurich
- Assessment
- Project (50 percent), written report (30 percent), final presentation (20 percent)
- Semesters
- Fall semester 2026
AI systems that pay off are also available as a talk or workshop for your leadership team. Request a talk