Data Science capstone projects batch #35

What happens when data science meets real-world challenges? Our latest Full-Time Data Science Batch #35 put their skills to the test with three ambitious final projects.

The projects highlighted the many ways data, AI, and creative problem-solving can open new possibilities.

See how our graduates are putting data science into practice and making an impact in the real world.
 

ProEngineers: Tender PDF-to-CSV Extractor for Construction Industry

Students: Vasili Areshka and Hanna Sliashynskaya

Construction companies receive multiple tender documents every day, and reviewing them is a slow manual process. Tenders are usually long, highly technical, and inconsistent in format, which makes processing difficult. In many cases, only a small subset of positions is actually relevant for a given company. Our client ProEngineers, who develop custom digital tools for the construction industry, challenged us to create a tool that specifically identifies facade and masonry-related positions. 

The goal was to build a solution that can automatically identify these positions in tender PDFs and extract them into a structured CSV output. To solve this, we chose a RAG approach, combining document retrieval with a large language model to find and extract the positions related to the target products. The graph shows the architecture of our solution: During the project, we experimented with different technical approaches to improve the quality of the results. This included testing different chunking strategies for splitting PDF text, refining keyword lists to better capture relevant products and descriptions, and iterating on prompts to guide the model more accurately. We also tested multiple language models and evaluated them using tailor-made metrics designed for this project.

For ProEngineers it is essential not to miss any relevant products from the tenders while keeping the extraction precise at cell level and minimising irrelevant rows. Our tests showed that the following models performed best at the task of extracting technical tender information:

The tool is ready to run on ProEngineers’ infrastructure and will be used in production.

Future plans include adding text recognition for image-based PDFs, improving chunking to reduce the amount of text sent to the LLM and even collecting the CSV outputs from our tool to train an independent model using machine learning techniques.
 

Smart Piece Counting: Computer Vision for Industrial Production Lines

Project by: Cristian Mourelle, Ivan Mateo Forcen, and Riyan Reyhan

Hadi-Plast GmbH is a German family-owned manufacturer based in Hövelhof, specializing in precision injection molding of thermoplastic components for the automotive, medical, and retail industries since 1977. Their production lines currently rely on high-precision industrial scales to count pieces on conveyor belts — an effective but expensive solution. As part of a capstone project with Hadi-Plast, our students developed a computer vision alternative: a system that detects, tracks, and counts individual pieces in real time as they move along the belt.

The challenge was far from trivial. The pieces are small, elongated plastic sticks that frequently overlap and pile up during transport, making both detection and tracking exceptionally difficult.
 

Custom Model Pipeline

The core of the project is a purpose-built pipeline combining deep learning detection with multi-object tracking. The team manually annotated 100 frames with Oriented Bounding Boxes (OBB) using Label Studio — a format that fits rotated and elongated sticks far more accurately than standard bounding boxes. This deliberate choice of manual annotation over automated labeling ensured high-quality training data, which proved essential given the small dataset size.

The annotated frames were used to fine-tune a YOLO11 nano OBB model over 300 epochs with early stopping. Data augmentation was carefully tuned for the fixed overhead camera setup: 180° rotation, mosaic, mixup, HSV jitter, and flips. The trained model feeds into ByteTrack, a multi-object tracker that assigns persistent IDs to each piece across frames and registers them as they cross a user-defined counting line. A recovery heuristic detects new IDs appearing past the line to handle pieces lost to ID switches, with spatial deduplication to prevent double-counting. The system outputs an annotated video with a live counter overlay and a CSV log with timestamps and positions for auditing.

The results are promising: a mAP50 of ~0.85 and 95% recall with zero false positives from background, all from just 100 labeled frames. Inference runs at ~35 fps on Apple Silicon, well above the 20 fps video rate, making real-time deployment viable. The main limitation lies in tracking under heavy occlusion, where identical overlapping pieces cause ID switches, a well-documented challenge even for state-of-the-art trackers (Zhang et al., ECCV 2022). Upgrading to BoT-SORT with visual re-identification would reduce these errors significantly (Aharon et al., 2022).

Importantly, the current test case represents one of the hardest scenarios for piece tracking. With larger, more separated, or visually distinct items, the exact same pipeline would achieve significantly higher accuracy, making it easily adaptable to different products and production line configurations across Hadi-Plast's operations.

 

Alternative Tools Exploration

Alongside the custom pipeline, the team explored Roboflow as a complementary approach. Roboflow offers an integrated platform for AI-assisted annotation, dataset management, and model training, providing a faster path from raw images to a working detector. This exploration demonstrated the range of available tools in the computer vision ecosystem and their potential for accelerating future iterations of the system.

Cristian, Ivan, and Riyan are proud to have contributed to Hadi-Plast's journey toward smarter, more efficient production monitoring.

 

Qumasoft: LLM Evaluation

Project by: Anil Jala, Gustavo Peredo & Hossam el din Baioumy
 

Introduction

The rapid rise of Large Language Models (LLMs) has left many companies balancing the need for innovation and the requirements of data privacy. Qumasoft, a German cybersecurity firm, actively addressed this challenge. 

Qumasoft provides a platform that supports product cybersecurity by helping companies manage regulatory requirements, streamline engineering efforts, and reduce reliance on external consulting. Its solution, quma.ai, combines automation and domain knowledge in product development, standards, and compliance to assist teams in documentation, reviews, and audit preparation while maintaining internal knowledge.

Data Sovereignty vs. External Dependency

For a cybersecurity specialist, prioritizing data integrity and sovereignty is a core requirement. For Qumasoft, this translates into a strategic approach that gives customers different options for deploying AI securely – from using powerful proprietary models to open-source LLMs that can be operated fully locally and self-hosted. Four strategic drivers are at the core of this approach:

• Greater data sovereignty: By running open-source LLMs locally, sensitive manufacturer and customer data can be processed entirely within the customer’s own IT environment. This provides greater control over data and how it is processed.
• Privacy based on customer requirements: A self-hosted open-source solution is particularly attractive for customers with stringent data protection and compliance requirements. It enables AI applications to be operated in a way that keeps sensitive data within the customer’s own infrastructure.
• Flexibility instead of dependency: Alongside closed models, Qumasoft offers open-source LLMs as an additional option. Depending on the use case, security requirements, and customer preferences, the most suitable model and deployment approach can be selected, without being tied to a single provider by default.
• Driving innovation: By enabling powerful AI to be operated securely and locally, Qumasoft positions itself as a pioneer in sovereign, self-hosted AI solutions for the European market.

From Manual Testing to Automated Auditing

Building on these strategic motivations, the team further explored relevant functionalities and approaches to support Qumasoft’s objectives. The focus was on evaluating and extending potential solutions in line with the company’s direction toward secure, efficient, and self-hosted AI systems. 

The methodology was designed to move from human-led verification to a scalable, automated pipeline.
 

1 - Manual Benchmarking:

Before automating the process, the team established a baseline through rigorous testing. We dispatched large-scale prompts (70k+ tokens) across diverse open-source and proprietary LLMs to test their "stamina" and precision.

 

2- The Multi-Agent Architecture

A Multi-Agent Flow was then created to replace manual oversight. This "LLM-as-a-Judge" architecture includes:

• Iteration Agent: Manages repetitive testing cycles to ensure a model's performance is consistent and not a one-off result.
• Judge Agent: Evaluates model outputs against strict dimensions, including Instruction Following, Knowledge Utilization, and Hallucination rates.
• Final Report Agent: Automatically transforms these specific metrics into a high-fidelity, branded PDF evaluation report for that specific model.

​​​​3 - Granular Evaluation and Benchmarking

The ultimate output was a system that generates a comprehensive, branded evaluation report for every model passed through the pipeline.

By analyzing these individual reports, we were able to create a definitive Benchmark. This bird's-eye view validated that Gemma 4 (31B) and Deepseek Reasoner achieved 100% relative accuracy, identifying them as the strongest candidates for Qumasoft's local deployment.
 

Conclusion

By implementing this automated auditing pipeline, Qumasoft is now equipped to select and deploy models based on objective metrics derived from detailed, model-specific reports. This ensures that as the AI landscape evolves, their commitment to cybersecurity and data sovereignty remains unshakeable.