Students

The rise of Large Language Models (LLMs) presents critical challenges related to hallucinations and the lack of updated parametric knowledge. To mitigate these limitations without incurring the high costs of retraining, this Bachelor’s Thesis designs, implements, and evaluates a decoupled architecture for Retrieval-Augmented Generation (RAG) systems. The technical justification lies in separating up-to-date textual information from the semantic search indices, thereby optimizing performance, latency, and computational cost efficiency. The methodology followed an iterative experimental approach using the SQuAD dataset (87,355 records). First, a local baseline was implemented by decoupling PostgreSQL (text storage) and ChromaDB (vector index). Building on this baseline, an exhaustive comparative study was carried out across 13 embedding models from the sentence-transformers library, measuring vectorization time, query latency, and accuracy via the Hit Rate metric. The generative cycle was then integrated with gpt-4o-mini and audited using the LLM-as-a-Judge technique with the GEval semantic rubric (Correctness). Finally, the system was migrated to a distributed cloud environment, evaluating a cloud architecture with Neon PostgreSQL and LanceDB persisted on Amazon S3 through quantized indexing (IVF-PQ). The empirical results showed that the Top-5 retrieval strategy optimizes contextual capture, achieving an HR@5 of 81.73% for contexts and 90.61% for titles with the lightweight all-MiniLM-L6-v2 model (average latency of 1.75 ms/query). In the generative evaluation, the system achieved a pass rate of 77.78% locally and 71.21% in the cloud. Likewise, the distributed deployment on AWS EC2 mitigated the residential network bottleneck by 63%, validating the scalability, resilience, and operational feasibility of the proposed infrastructure.

Retrieval-augmented generation, or RAG, answers questions from passages retrieved out of a corpus, and every claim can cite where it came from. Such a system is easy to start and hard to sustain: held together by scripts it works, but changing one piece ends up touching them all. The goal of this work is to implement a modular RAG library over the Open Web Index that runs on a single local machine and is ready to scale out by replacing components, and to evaluate its retrieval stage systematically. The library covers the whole cycle, from ingest to answer, with seven components that configuration can substitute under an orchestrator holding no logic of its own. It offers two retrieval modes: sequential, over a single collection, and collective, which splits the index into partitions searched separately and merges their results into one ordered list. Systematic evaluation covers retrieval: thirty hand-annotated queries, k = 10, over a real corpus and a campaign of 1.36 million vectors. The correctness of the collective mode was checked against an exact bruteforce reference rather than against the sequential mode: the merge introduces no error, and the partitioned path proves at least as faithful as the single collection. On one machine, however, it never becomes faster: coordinating the partitions costs more than it saves, and the workers contend for the same resources. What is measured is that searching one isolated partition costs 14 to 16 milliseconds against the 74 of the whole collection: the gain is real, and the design leaves ready, behind the same interfaces, the spread across machines with their own memory.

This document presents CAPPM (Conversational Agents Platform with Pluggable Models), a platform designed to drive the development and testing of AI applications with total abstraction of the underlying models. CAPPM allows for the easy creation of intelligent agents that can be configured from custom documents, utilizing Retrieval-Augmented Generation (RAG) techniques to answer questions and solve tasks within specific information domains. The solution stands out for its flexibility in adjusting agent behavior and relies on an elastic and scalable cloud architecture that ensures maximum efficiency and adaptability in multi-user environments. In addition to detailing the architecture and providing a user guide, the document includes a system evaluation in a real-world environment. The analysis gathers user feedback, evaluates the models, identifies limitations, and outlines lessons learned along with proposals for future improvements.

The Cloud Computing paradigm and, in particular, serverless architecture, offers substantial advantages such as scale-to-zero and a strict pay-as-you-go model. However, most current vector databases rely on always-on instances, which guarantee low latency but drastically increase operational costs and reduce resource efficiency. This Bachelor’s Thesis aims to design and implement a fully serverless and dynamic vector database capable of supporting continuous data ingestion while remaining highly competitive with state-of-the-art solutions. To achieve this goal, the project builds upon a static baseline architecture provided by the CloudLab research group (URV). From this starting point, an iterative development methodology was applied where, through empirical analysis of an initial implementation, an optimized architecture was designed based on asynchronous state compaction and an innovative hybrid query model. The results demonstrate that the proposed architecture achieves superior performance, both in execution time and operational costs, compared to other commercial market options when subjected to massive analytical workloads (ingestion operations and grouped semantic queries or batch queries). Finally, it is concluded that, despite the intrinsic orchestration overhead in low-frequency use cases or isolated requests, the system represents a structurally robust and affordable solution to democratize large-scale vector storage.

This Final Degree Project presents the design and development of an automated auditing solution focused on identifying underutilized resources in hybrid environments based on VMware vCenter infrastructures and Amazon Web Services (AWS). The main objective of the project is to simplify the detection of inefficient resource allocation and optimization opportunities while reducing the manual effort required for infrastructure audits. The solution uses Ansible for orchestration and Python for data processing, integrating with VMware and AWS APIs to collect information from virtual machines, ESXi hosts, datastores, EC2 instances, EBS volumes and related resources. Several analysis rules are applied to detect oversized virtual machines, long-term inactive resources, duplicated IP addresses, workload imbalances and underutilized cloud services. The generated results are automatically consolidated into detailed reports and executive summaries delivered through Microsoft Teams. The validation process confirmed the correct operation of the solution and revealed multiple optimization opportunities across both VMware and AWS environments. The project required significant design, development, integration and testing efforts, combining automation, virtualization and cloud technologies into a centralized infrastructure auditing solution.

AI inference services deployed in cloud-based environments face a dual challenge of meeting latency-related SLO requirements and maximizing cost and resource efficiency. This thesis experimentally compares two horizontal pod autoscaling approaches for AI inference on Kubernetes: the CPU-based Kubernetes Horizontal Pod Autoscaler (HPA) and the concurrency-based Knative Pod Autoscaler (KPA). The evaluation was performed on Google Kubernetes Engine using two CPU-only AI inference workloads with different CPU-intensity levels: TinyBERT for sentiment analysis (low CPU footprint) and YOLOv12n for object detection (high CPU footprint). Both autoscalers were evaluated using simulated traffic patterns, including steady load, increasing load, bursty traffic and intermittent traffic with idle periods of variable length. Knative workloads were evaluated with both warm-start and cold-start initial states. The study considered metrics such as average and tail latencies, request failure rate, replica counts, concurrency, CPU and memory usage and allocation, and estimated cost efficiency. The results show that there is no single autoscaler that performs best in all situations. Knative can provide lower resource allocation and higher cost efficiency when idle periods are sufficiently long and exceed the workload’s scale to zero time, while HPA offers lower latency during bursty traffic and after idle periods in which Knative scales to zero. The first hypothesis evaluated by the study involved an idle period threshold, at and above which HPA has higher resource allocation and lower cost efficiency than Knative, and this threshold was confirmed for both workloads at an idle period of 3.0 times the scale to zero time. The second hypothesis involved an idle period threshold, at and above which HPA has lower latency than Knative, and this threshold was confirmed within the considered idle period range at 0.8 times the scale to zero time. The findings show that the choice of autoscaler for AI inference should consider the traffic pattern, workload CPU-intensity level, the presence of cold starts and specific service requirements regarding performance and cost efficiency.

This Master’s Thesis presents a rigorous comparative analysis of three representative automotive production philosophies United States, Europe, and Japan using Discrete Event Simulation (DES) combined with a multi-objective optimization approach. The study is based on the development of equivalent simulation models that share the same product, process sequence, and input conditions, allowing the impact of production philosophy on system performance to be isolated. In a first stage, the baseline behavior of each model is analyzed without optimization, using key performance indicators such as throughput, work-in-process (WIP), lead time, and rework rate. Subsequently, saturation and breakdown scenarios are introduced to evaluate the structural robustness of each system under operational disturbances. In a final stage, a multi-objective optimization process is carried out through realistic structural modifications, mainly by adjusting parallelism at critical stations, while keeping processing times and arrival conditions unchanged. The results show that no universal optimal configuration exists; instead, each system exhibits inherent trade-offs between throughput, WIP, and lead time that strongly depend on the underlying production philosophy. Pareto front analysis is used to visualize these trade-offs and to assess how different philosophies respond to equivalent optimization actions. The thesis concludes that the integration of discrete event simulation with multi-objective optimization is a powerful framework for analyzing complex production systems and for deriving insights applicable to real industrial decision-making contexts.