Data Science capstone projects batch #34

We’re thrilled to celebrate the achievements of our latest graduates from Zurich’s Full-Time DS Batch #34, who have just wrapped up their Data Science journey with 5 remarkable, real-world projects. 

This round of final presentations showcased how data science and AI can drive tangible impact across industries, from transforming business development workflows to reinventing the way of the new market discovery. 

Take a look at how our graduates are using data science to generate insights, push boundaries, and create real-world impact. 

Transforming Raw Business Ideas Into Real-Time Intelligence  

Students: Mohammed, Lamees, Amy, Ashley  

Galaxon is an AI-powered business intelligence assistant built to help innovators move faster from vague ideas to clear, actionable insights. Today, innovation is often slowed by inefficient research cycles. Developing a new business idea can take six to eight months, many concepts start out loosely defined, and traditional reporting processes are lengthy and difficult to interpret.  

 

The team is set out to solve this problem by creating a system that transforms rough business ideas into structured snapshots using AI-driven reasoning. Instead of relying on static data sources, Galaxon performs a real-time market scan, ensuring that insights reflect the current state of the market rather than outdated information.  

  

 

The overall system design is illustrated in the first image above, which shows Galaxon’s architecture. This architecture enables Galaxon to remain flexible, scalable, and responsive to fast-changing business environments.  

 

The result of this process is a concise and structured insight report, as shown in the second image below, which presents an output summary generated by Galaxon. This summary highlights the most relevant findings and key insights, allowing users to focus on decision-making rather than information overload.  

 

The benefits of Galaxon are clear. It significantly reduces research time, improves clarity in early-stage idea development, and helps innovators make informed decisions faster. By replacing long research cycles with real-time intelligence, Galaxon empowers teams to move confidently from ideation to action.  

For future development, the team plans to enhance the user interface to improve usability, interpretability, and overall interaction efficiency. Subsequently, they will refine the core system through advanced prompt engineering techniques to strengthen reasoning capabilities and increase the precision and consistency of generated insights. Finally, the aim is to reduce system latency to enable faster response times and support a more seamless real-time user experience.  
 

Identifying High‑Potential Segments for Transaction Growth  

Students: Olga Bazhenova, Jaime Antolin, Glory Mary Givi, Mudathir Awadaljeed 

SIX Group, a key provider of Switzerland’s financial market infrastructure, challenged our team to identify where the company has the highest untapped potential for transaction growth. The goal was to move beyond intuition and build a datadriven view of sectors, merchants, and regions that offer the strongest opportunities for activation and reactivation.  

To do this, Olga, Jaime, Glory, and Mudathir combined aggregated, anonymized internal transaction data with public demographic statistics, regional indicators, and company registry information from ZEFIX. This allowed us to build a comprehensive picture of the Swiss merchant landscape and quantify where growth is most achievable.  

Their analysis revealed three major growth levers:  

1. Newcomers to Switzerland with immediate billing needs — With over 125000 new foreign residents arriving in 2025 — highly concentrated in Zurich, Vaud, Geneva, and Bern — this segment represents a predictable and scalable source of new transactions. Newcomers have immediate billing needs and high digital readiness, making them ideal early adopters.  

2. Merchant Gaps and Reactivation Potential — Among more than 5000 merchants, they identified:  

• 387 new merchants who joined in 2025 but have zero transaction activity  

• A large pool of inactive but still operating companies, representing a strong reactivation opportunity  

• A small group of merchants responsible for over 50% of the total transaction gap  

These insights highlight where targeted outreach and onboarding support can unlock significant volume.  

3. HighPotential Regions — Using a composite score that blends customer density and adoption gaps, we mapped the regions with the strongest growth potential. Clients located in these highpriority areas represent over 200,000 additional transactions if they reach expected performance levels. This regional view helps SIX focus activation efforts where they will have the greatest impact.  

 

 

Based on these insights, the team developed targeted recommendations for SIX’s marketing and business development teams, including:  

 

• Activation strategies for newcomers  

• Reactivation campaigns for inactive merchants  

• Sectorspecific approaches for belowmedian performers  

• Regionbased prioritization for efficient resource allocation  

Together, these strategies form a unified framework for unlocking transactional growth across Switzerland.  

 

How Multi-Agent AI Improves Complex Decision-Making Through Structured Debate  

Students: Alexey Vitsenko,  Martin Oehmke,  Rizwan Nasir  

Every major business decision — from engineering a new product to reshaping company policy — carries significant risk. The challenge is not only choosing the right option but ensuring that all relevant perspectives have been considered. Cognitive bias, unchallenged assumptions, and limited viewpoints can quietly shape outcomes, even in highly experienced teams. 

This challenge formed the foundation of this project developed in collaboration with BMW. The student team explored how artificial intelligence could be used not to deliver a single “correct” answer, but to rigorously challenge decisions before they are made. Their solution investigates a multi-agent AI architecture in which specialized agents debate, critique, and refine ideas — simulating structured disagreement to improve decision quality.  

From Answers to Arguments: The Core Concept 

Traditional AI systems typically respond to a user query with one consolidated answer. In contrast, the team designed a multi-agent system where multiple AI agents approach the same question from different expert perspectives. 

Each agent is assigned a clearly defined role within the decision process: 

• Proposer – formulates an initial solution or recommendation 

• Critic – challenges assumptions, identifies weaknesses, and stress-tests logic 

• Refiner – improves proposals by incorporating valid critiques 

• Reflector – evaluates the reasoning process and highlights remaining uncertainties 

Rather than operating independently, these agents engage in a structured, rule-based debate. The goal is not consensus by default, but informed synthesis. 

The workflow follows a five-step reasoning cycle: 

1. Query & Context Definition – the problem is framed and relevant agents are selected 

2. Initial Proposals – agents generate independent viewpoints 

3. Structured Debate – agents challenge each other’s assumptions and evidence 

4. Iterative Refinement – arguments evolve through multiple critique rounds 

5. Synthesized Outcome – insights are consolidated into a decision-ready summary 

This approach transforms AI from a question-answering tool into an analytical partner that exposes trade-offs and uncertainty. 

Orchestrating the Debate: The Central Coordinator - To ensure structure and coherence, the system relies on a Central Coordinator. Rather than contributing opinions, this component manages the reasoning process itself. The coordinator determines which agents participate, tracks argument evolution, enforces debate rules, and ensures balanced representation of perspectives. Once the debate converges, it synthesizes validated inputs into a single, coherent output designed for human decision-makers. 

Grounding Reasoning with Knowledge Base and RAG - For real-world applicability, the team emphasized that AI debate must be grounded in company-specific knowledge rather than generic internet data. To address this, the system integrates Retrieval-Augmented Generation (RAG). 

Beyond Engineering: Applying the System to HR Strategy - While automotive engineering is a natural use case, the project demonstrates broader applicability. The team illustrated this with an HR strategy example: Should a company introduce a four-day work week without reducing salaries? 

Instead of producing a single recommendation, the system initiates a debate between specialized agents: 

• HR Strategy Agent – evaluates employee motivation, retention, and employer branding 

• Finance Agent – analyzes productivity impact, cost structures, and financial risk 

• Operations Agent – assesses feasibility, delivery timelines, and coordination challenges 

The final output is not a simple yes or no, but a structured recommendation that balances people impact, financial risk, and operational feasibility — potentially including phased rollouts or conditional safeguards. 

Reducing Bias and Increasing Decision Confidence - A key outcome of the project is the systematic reduction of bias. By forcing multiple perspectives to challenge each other, blind spots and implicit assumptions become visible early in the process. This is particularly critical in safety-sensitive environments, where overlooked factors can have serious consequences. 

Additional benefits include: 

• Consistent, high-quality reasoning across departments 

• Transparent audit trails of how decisions were formed 

• Faster management decisions through upfront structured analysis 

AI as a Framework for Disciplined Reasoning - This capstone project shows how multi-agent AI can improve complex decision-making by structuring disagreement rather than producing a single “best” answer. By making assumptions explicit and surfacing risks and trade-offs, the system enhances human judgment instead of replacing it. As enterprise AI matures, such approaches highlight AI’s role as a transparent and scalable framework for disciplined reasoning alongside human expertise. 
 

Lola - HR Contract Management System  

Students: Alejandra Vicaria, Catalina Cimpanu, Dora Novoa, Laura Fatta  

The average HR department spends countless hours on manual data entry and complex paperwork. These repetitive tasks are often prone to errors, such as typos or inconsistencies, leading to frustration and inefficiency. In an era where AI can streamline workflows, HR teams have the potential to reclaim this lost time. This empowers them to focus on strategic people management, but only if they have the necessary intelligent tools to do so.  

The team was tasked with doing precisely this for BAK Economics. BAK Economics is a leading independent economic research institute and consultancy, providing empirical economic studies, in-depth analyses, and model-based forecasts for businesses, public institutions, and policymakers in Switzerland and internationally. To support their internal operations, BAK challenged the team to replace manual form-filling with an intelligent workflow that could automate the creation and management of employment contracts.  

Given the complexity of managing multiple contract versions and the tedious nature of manual entry, the project would be considered a success if the team managed to:  

• Build a conversational interface to guide users through the process.  

• Automate the selection of the correct contract type based on user input.  

• Generate error-free, ready-to-sign documents (PDF & DOCX).  

The team implemented a strategy that leveraged AI and modern web frameworks to exceed these expectations. This resulted in:  

• 5 Distinct Contract Types successfully automated (Permanent, Internship, Hourly, Amendment, and Conversion.  

• Intelligent Features such as natural language date parsing (e.g., understanding "next Monday") and automated typo handling.  

• Multi-language Support for the interface (German, English, Spanish).  

• One Comprehensive Web Application that handles the entire lifecycle, managing not just the conversation but also employee records, database entries, and final document delivery.  

In order to achieve these results, the team leveraged a variety of tools including Large Language Models (LLMs), Python-based document processing libraries, and cloud infrastructure.  

The first component involved creating the "Brain" of the system: The AI Agent. This layer processes user messages, manages the conversation state, and validates collected data.  It includes a smart correction workflow, allowing users to review and modify details before finalization. It automatically routes the user to the correct contract "path" (e.g., New Hire vs. Existing Employee) without the user needing to understand internal naming conventions.   

The second component focused on Document Generation and Data Management. Once the agent collects the necessary information (e.g., salary, workload) the system uses python tools to fill templates. To further streamline the process, the system utilizes search algorithms, allowing the AI to correctly identify existing employees in the database even if the user types a partial name or makes a typo. All data, including audit logs and employee records, is securely stored in a PostgreSQL database, ensuring data integrity and allowing for easy "Employee Lookup" for future contract changes. The results can be seen in the logic flow below, which maps how the system handles different employment scenarios.  

 

The third component centered on a comprehensive web application and database architecture that extends far beyond a simple chat interface. The team built a robust full-stack platform, using Flask and deployed it on AWS, backed by a relational database designed to manage the entire HR lifecycle securely. This architecture allows administrators to manage core data directly within the app (e.g., employee records, job positions, and departments). Beyond just generating documents, the system tracks contract versions, logs salary and workload change requests, and maintains a detailed audit trail of every user action. Additionally, the system supports bulk data operations, allowing administrators to generate Excel exports for reporting alongside individual PDF contract generation. This ensures that while the AI agent drives the conversation, the underlying application guarantees data integrity, security, and full administrative oversight.   

 

Through these steps, the team transformed a manual, error-prone task into a streamlined, AI-powered workflow. The tool not only systematically produces clean, high-quality contracts but also provides BAK Economics with a scalable foundation for future HR automation.   
 

Tygo Telecom AI Chat Assistant 

Students: Fares Alharbi, Anas Alsaedi, Deena Alkhlewi, Ahmed Althiabi 
 

 Tygo is a Saudi mobile telecommunications provider offering flexible connectivity and customer-centric data plans across the world. In this capstone, the system was built for Tygo to address challenges in retrieving and reasoning over large volumes of unstructured data. 

AI models have become powerful tools for modern businesses. For relatively simple use cases, off-the-shelf solutions can often be implemented quickly and deliver immediate value. However, as projects grow in scale and complexity, the limitations of generic AI solutions become apparent. Large, long-running initiatives introduce challenges such as unstructured legacy data, inconsistent formats, and increased risk of hallucinations — especially when models are overloaded with instructions in an attempt to “understand” decades of accumulated context. Some AI models still don't know how to spell raspberry (it's quiet funny look it up!) 

So how can an AI solution be made scalable and stable while still remaining highly helpful and responsive through well-designed prompting techniques? For example, in a digital transformation project involving thousands of files dating back to the 1970s, the data can exist in many different formats and structures, making it nearly impossible to store in a fully standardized way. Retrieving such information later often require painful manual processes — an area where AI could provide significant acceleration. The challenge, however, is how to implement an AI solution capable of handling projects of this scale while preventing hallucinations, especially when the model must be instructed about decades of accumulated complexity. 

That's what Fares, Anas, Deena, and Ahmed have achieved with Tygo by building a chatbot that retrieves data plans for customers in a complicated market:  

 

  

 

They studied Tygo business and decided on avoiding storing everything in a raw text format for the AI model to retrieve later from the vector database, but to instead add metadata alongside the embedding data so the database could be sorted and filtered before the AI model get to it in the first place, that way this allows one to do two important things: 

• Reduce the amount of data needed to deal with at the query level  

• Handle prompting dynamically with the metadata related filters for a given request.  

For example, imagine you have 100 items to sell. Instead of allowing the AI to search across all items to identify the cheapest 10%, prices can be stored as metadata alongside the embeddings and used to sort or filter results in advance. This way, the language model only needs to reason over the most relevant subset. When scaled from hundreds to thousands or even millions of items, this approach delivers substantial efficiency gains. Additional metadata such as data source, pricing, or document dates can be included to keep large projects performant and manageable. 

 

To implement this approach, data is collected regardless of its original format or source. Documents are chunked, embedded, and enriched with metadata such as source identifiers, dates, pricing, or other sortable attributes (see the diagram above). These enriched chunks are then stored in a vector database, forming the foundation of the Tygo chatbot architecture. 

This project demonstrates how thoughtful data structuring and retrieval strategies can significantly improve the reliability and scalability of AI-powered systems in complex enterprise environments. 
 

Conclusion 

With the completion of the Data Science capstone projects of the Full-Time Batch #34, we celebrate the impressive achievements of our graduates. Their work clearly demonstrates what happens when technical expertise meets creativity: data science unlocks its full potential for innovation. 

A huge thank you goes to our partners, mentors, and instructors who supported the teams throughout their journey. Your knowledge, guidance, and collaboration helped transform ideas into working, real-world prototypes. 

To our graduates: your curiosity, perseverance, and ambition to solve real problems are what make this community so special. We’re excited to see where your path leads next - and how you will continue shaping the future of AI and data-driven innovation. 

If these stories inspire you, learn here how you can become part of the next Data Science batch at Constructor Nexademy and start your own innovative project.