Hi, I'm Alvaro

Project Overview

This project was developed as a collaborative research initiative involving ETH Zurich, HSLU, and Google DeepMind. The goal was to design an end-to-end Retrieval-Augmented Generation (RAG) assistant that answers complex, multilingual questions about ETH’s news archive while citing primary sources.

The Problem

News archives feature long-form multilingual (German and English) articles where strictly lexical search underperforms: keyword-only queries miss paraphrased facts, while naïve dense retrieval over-indexes semantically similar but contextually irrelevant passages. We needed a pipeline that (1) understands nuanced questions, (2) surfaces the exact supporting snippets, and (3) synthesizes grounded, audit-ready answers.

Technical Approach

1. Data Engineering & Chunking

• Parsed ETH news HTML into structured records, preserving language metadata and publication dates.
• Applied hybrid chunking (Recursive Character + semantic boundaries) to keep discourse units intact.
• Built a “gold set” of 25 benchmark questions with GPT-4o-generated relevance annotations to drive evaluation later.

2. Hybrid Retrieval & Research Agents

• Dense backbone: multilingual distiluse-base-multilingual-cased-v1 embeddings indexed in FAISS.
• HyDE expansion: GPT-4o generates three hypothetical documents per query; their averaged embeddings boost recall on under-specified questions.
• Combined standard dense scores and HyDE scores via weighted normalization to produce a 100-candidate pool per query.

3. Post-Retrieval Refinement & Fusion

We benchmarked seven rerankers, each optimized for a different trade-off:

1. EcoRank (MiniLM→Electra cascade) for budget-aware precision.
2. FlagReranker v2-m3 (FP16) for multilingual robustness.
3. Rankify-style BGE reranker implemented directly with Hugging Face for controllable batching.
4. Cohere Rerank v3.5 offloads heavy lifting to an API, providing consistent latency on weaker GPUs.
5. Qwen3-Reranker-4B for LLM-grade judgments where latency is secondary.
6. Multislot diversity reranker (MMR-style) to avoid redundant chunks.
7. Reciprocal Rank Fusion (RRF) across the strongest trio (EcoRank, Flag, Cohere) with async execution—the fused list consistently achieved the highest Precision@10 with latency limited by the slowest component.

4. Guided Response Synthesis

• Retrieved contexts are serialized to JSON (chunk text, source ID, publication date) and fed to GPT-4o with a system policy enforcing neutrality, concision, and JSON-only answers.
• Model outputs both the final answer and the chunk IDs it relied on, ensuring transparent traceability.

5. Comprehensive Evaluation

• Automated metrics: semantic Exact Match (MiniLM embeddings), semantic F1 (token-level cosine matching), BLEU, ROUGE-1/2/L.
• Human review: 1–5 ratings for relevance, correctness, clarity across all 25 questions.
• Findings: Fusion reranking lifted Precision@10 to the mid-0.7 range while EcoRank delivered the best MRR at a fraction of the latency. Human evaluators gave average scores of 4.4 (relevance), 3.7 (correctness), 4.9 (clarity); correctness dips aligned with questions whose retrieval pool lacked high-scoring ground-truth chunks.

Project Phases

Phase 1 – Multilingual Data Engineering
Established the benchmark dataset, chunk repository, and relevance annotations to ensure repeatable experimentation.

Phase 2 – Hybrid Retrieval & Research Agents
Implemented dense+HyDE retrieval with on-the-fly query rewriting; stored 100-candidate pools per query for deterministic downstream testing.

Phase 3 – Post-Retrieval Refinement & Response Synthesis
Evaluated multiple rerankers, introduced asynchronous RRF fusion, and handed the top-10 ranked chunks to GPT-4o with strict JSON prompts for grounded answer generation.

Phase 4 – Evaluation & Insight Extraction
Compared machine metrics to human ratings, visualized metric heatmaps per question, and correlated failure cases back to retrieval coverage to guide future data augmentation.

Results

• Fusion of EcoRank + FlagReranker + Cohere improved factual hit rate on challenging “who/when” questions versus any single reranker.
• Automated semantic F1 scores mirrored human correctness ratings except for intentionally open-ended queries, highlighting where quantitative metrics under-credit acceptable paraphrases.
• The pipeline answered 20 of 25 benchmark questions correctly (per human judgment) and gracefully declined when evidence was missing, producing source-linked, portfolio-ready responses.

Project Overview

This project was developed as a core initiative under the BYDO project, a collaboration between ON AG and HSLU. The objective was to automate the assignment of Harmonized System (HS) Codes for international imports and exports. By leveraging product descriptions and manufacturing specifications, the system identifies the precise regulatory codes required to ensure trade compliance and minimize manual overhead.

The Problem

Classifying products for international trade is a notorious logistical bottleneck. While the first six digits of HS codes are standardized globally, individual countries apply unique suffixes, formatting rules, and local interpretations. For a global brand like ON AG, manual classification is slow and prone to human error. Traditional keyword searches often fail because regulatory language is highly specific and does not always align with commercial product descriptions. The goal was to build a pipeline that could ingest complex manufacturing data and output accurate, audit-ready, country-specific codes.

Technical Approach

1. Data Engineering & Standardization

The foundation of the project relied on the Google Cloud Stack to handle large-scale regulatory and product datasets.

• Data Sourcing: Raw product specifications and historical classification records were centralized in BigQuery.
• Cleaning & Normalization: Python-based scripts standardized product descriptions and material compositions.
• Global Mapping: A unified schema was developed to manage the variations in HS code formats across different jurisdictions, ensuring the model could generalize across various international markets.

2. Model Benchmarking & Selection

Three distinct modeling approaches were evaluated to determine the best balance between precision and contextual understanding:

• Gradient Boosted Decision Trees (GBDT): Utilized as a baseline to process structured features from the manufacturing data.
• Multi-Layer Perceptron (MLP): A neural network approach designed to capture non-linear relationships within the feature set.
• Google AI Gemini (via Vertex AI): An LLM-based approach utilizing Context Engineering. This method allowed the model to process nuanced text descriptions and cross-reference them with regulatory definitions provided in the prompt context.

3. Context Engineering & Retrieval

While traditional ML models performed well on structured data, Gemini demonstrated superior performance in interpreting complex product descriptions. Through Vertex AI, specific context, such as material composition, footwear components, and product intent, was engineered into the prompts. This allowed the model to reason through classification rules rather than simply performing pattern matching.

4. Human-in-the-Loop Verification

To ensure compliance and build trust with trade specialists, the system was designed with a “source-first” philosophy. For every HS code suggested, the chatbot retrieves and displays the top 3 most relevant sources (such as specific regulatory clauses or historical classification precedents). This allows human experts to instantly verify the model’s logic and ensures the output is audit-ready.

5. Prototyping

A functional prototype was built using Streamlit to bridge the gap between the data pipeline and the end-user.

• Document Selection: Users can choose relevant technical documents or manufacturing specs to guide the query.
• Interactive Chatbot: A conversational interface allows users to interrogate the classification, asking for clarification on specific rules or material-based distinctions.

Project Phases

Phase 1 – Data Gathering & Standardization Consolidated global trade data into BigQuery. Addressed the challenge of differing code lengths and formats by creating a standardized preprocessing pipeline that aligns country-specific codes to a core product identity.

Phase 2 – Model Development & Evaluation Developed and trained GBDT and MLP models to establish a performance baseline. Simultaneously, established a testing framework for LLM prompts using Vertex AI to compare traditional ML performance against generative AI capabilities.

Phase 3 – Context Engineering & Tooling Refined the Gemini implementation by providing high-quality reference data. Developed the Streamlit interface to transition the project from a backend model to a functional internal tool capable of citing its sources.

Results

• Higher Accuracy on Complex Items: The Gemini-based approach, supported by context engineering, significantly outperformed traditional models on products with ambiguous material compositions.
• Audit-Ready Transparency: By providing the top 3 relevant sources for every answer, the system eliminated the “black box” nature of AI, allowing for rapid human verification.
• Enterprise Scalability: Utilizing the Google Cloud ecosystem (BigQuery and Vertex AI) ensures the pipeline can handle the increasing volume of ON AG’s global product catalog while maintaining low latency.

Research Overview

Developed in collaboration with Mammut Sports Group AG and the Lucerne University of Applied Sciences and Arts, this Master’s Thesis explored the cutting edge of textile engineering. The research focused on Aramid-reinforced hybrid ropes, a revolutionary technology designed to prevent the most catastrophic failure in climbing: rope severance over sharp rock edges.

The project bridged the gap between physical laboratory experiments and digital predictive science, using data to quantify exactly how Aramid fibers change the “safety profile” of a climber’s most critical lifeline.

The High-Stakes Problem: Sharp Edge Failure

Standard climbing ropes are made of Polyamide (Nylon), which provides a “soft catch” but is vulnerable to shear forces. While integrated Aramid (the material in bulletproof vests) can stop a blade, its extreme stiffness raises concerns:

• The Shock Load: Does it transmit too much force to the climber?
• The Aging Factor: Do the coarse aramid fibers wear the soft nylon core from the inside during repeated use?
• The Conditional Advantage: How does the Aramid fibers improve the rope cut resistance under different angles and tension?

Methodology: From Drop Towers to Algorithms

The research utilized a three-pillar methodology to dissect rope behavior across nine distinct models and hundreds of controlled tests.

1. Laboratory Stress Testing

Conducted high-precision experiments on the Mammut Drop Tower and proprietary Edelrid Cut-Resistance machines:

• Single Drop Tests: Measuring peak impact force and dynamic elongation.
• Endurance Cycling: 30 repeated drops per specimen to measure mechanical aging.
• Sharpe Edge Simulation: Testing severance under variable blade angles (120° vs 160°) and extreme static loads (up to 300kg).

2. Statistical Inference (GEE)

Leveraged Generalized Estimating Equations (GEE) with Cluster Wild Bootstrap to account for dependencies within rope models and manufacturing batches. This ensured that findings were statistically valid despite the complex, hierarchical nature of the datasets.

3. Predictive Modeling (Machine Learning)

Built and benchmarked regression architectures to predict dynamic performance:

• Linear Baselines: Elastic Net for feature selection.
• Ensemble Methods: Gradient Boosting for non-linear threshold identification.
• Neural Networks (MLP): Modeled the continuous physics of polymer stretching, achieving >97% accuracy in predicting peak impact forces.

Key Insights & Results

The research provided a definitive, data-backed answer to the viability of hybrid rope technology:

The “Hard Catch” Trade-off

Aramid-reinforced ropes are notably stiffer than polyamide controls, producing higher peak arrest forces and lower dynamic stretch. They stay within certification limits but deliver a firmer catch that loads the protection chain more aggressively.

Conditional Superiority in Cut Resistance

Under moderate edge angles and loads, aramid constructions hold a clear advantage in resisting sharp-edge failures. Once angles tighten and tension rises, that lead rapidly narrows, indicating extreme geometries can negate the material benefit.

Parallel Degradation

Across thirty-drop endurance sequences, both constructions stiffened at essentially the same rate. No evidence surfaced of accelerated internal abrasion in the hybrid build, supporting parallel aging behavior.

Virtual Prototyping

Machine-learning surrogates reproduced the main laboratory trends well enough to vet new designs digitally, pointing to shorter R&D loops with fewer physical prototypes.

Impact

This thesis provided Mammut with the empirical evidence needed to position their “Core Protect” technology in the alpine market. It defined the “safe operating envelope” for cut-resistant ropes and demonstrated how data science can accelerate material innovation in the sports equipment industry.

Project Overview

This project was developed as part of the Customer Data Analytics module at HSLU in collaboration with the FinTech platform YAPPEAL. The objective was to transform raw card transaction data (2021–2023) into actionable business intelligence by identifying customer segments and quantifying churn risks through statistical modeling and interactive visualization.

The Problem

Financial institutions possess vast amounts of transactional data, yet extracting meaningful behavioral insights is challenged by high dimensionality and a mix of categorical (Merchant Codes, Countries) and numerical (Amount CHF) features. The goal was to build a system that identifies high-value segments and detects “silent churners”—customers whose spending activity is significantly declining over time.

Technical Approach

1. Data Engineering & Enrichment

• Multi-Source ETL: Merged transaction records from 2021, 2022, and 2023 while handling missing MCC (Merchant Category Code) metadata.
• Contextual Enrichment: Integrated Swiss public and “other” holidays (e.g., Black Friday, Valentine’s Day) to evaluate seasonal spending spikes.
• Categorical Mapping: Developed custom hierarchy logic to map 500+ unique MCCs into 15 macro-categories and specific subcategories like “Airlines” and “Public Transport.”

2. Advanced Preprocessing

• Outlier Detection: Applied log-transformation to amount_chf to handle the heavy-tailed distribution of transaction values, using IQR-based bounds to isolate significant outliers.
• Time-Series Extraction: Derived trx_timing (Morning, Afternoon, Evening, Night) and weekday features to build temporal spending profiles.

3. Behavioral Clustering

• K-Modes Implementation: Used the K-Modes algorithm to cluster categorical features (Airlines, Continents, Weekdays) without losing information to dummy-variable sparsity.
• Dimensionality Reduction: Utilized PCA (Principal Component Analysis) and t-SNE to project high-dimensional clusters into 3D space for evaluation.
• Segment Discovery: Identified distinct personas, such as “European Budget Travelers” (high Ryanair/EasyJet frequency) and “Local Grocery Shoppers.”

4. Churn Prediction & Retention Analysis

• Retention Slope: Calculated a quarterly “Slope” (linear regression coefficients) for every customer by tracking transaction frequency from 2021-Q1 to 2023-Q3.
• Risk Assessment: Customers with negative slopes above a specific threshold were flagged as “at-risk,” enabling targeted loyalty interventions.

5. Tableau Visualization

• Built integrated Tableau dashboards featuring:
  • Geographic Heatmaps: Spending density by ISO country codes.
  • Churn Storyboards: Interactive views of retention trends across different customer clusters.
  • Ratio Analytics: Visualizing User-to-Transaction ratios to differentiate between high-frequency and high-value users.

Project Phases

Phase 1 – Data Foundation & Preprocessing
established a clean dataset with unified currency (CHF) and enriched holiday metadata.

Phase 2 – Feature Engineering & Profiling
Mapped counterparts to specific airlines/continents and introduced temporal features for deeper behavioral context.

Phase 3 – Unsupervised Machine Learning
Executed K-Modes and K-Means experiments to segment the audience, optimized using Silhouette scores and Inertia plots.

Phase 4 – Dashboarding & Strategy
Translated statistical clusters into Tableau visualizations to provide business-ready insights into customer churn and market potential.

Results

• Successfully segmented a customer base of thousands into 4 actionable behavioral clusters.
• Identified the “Slope” metric as a reliable early-warning indicator for customer churn, highlighting significant findings for 2023-Q3 period.
• Delivered a comprehensive Tableau workbook providing stakeholders with deep-drills into category-specific spending behavior and geographic trends.

Professional Overview

This role was based within the Quality Management department at Mammut Sports Group AG, focusing on the modernization of data pipelines and the rigorous testing of technical mountain sports equipment. The position involved bridging the gap between physical laboratory testing and digital data architecture, ensuring that product safety and manufacturing standards were met across global supply chains.

Data Infrastructure & ERP Migration

A primary objective was maintaining data integrity during a large-scale transition of the company’s digital backbone.

• ERP Migration: Managed the update of the quality database during the migration from a legacy system to Microsoft Dynamics 365. This included reconfiguring standalone programs and files that relied on specific database entries to ensure zero data loss.
• Database Management: Leveraged Microsoft SQL and Excel to clean and pipeline testing data from laboratory servers to central storage, ensuring that historical records remained accessible and standardized.
• Documentation: Centralized all technical SOPs, testing protocols, and migration logs on SharePoint, creating a unified “source of truth” for the global quality team.

Automation & Strategic Reporting

To improve operational speed, several manual reporting processes were replaced with automated workflows within the Microsoft Stack:

• Provider Reporting: Automated the generation and distribution of quality reports to external providers. This replaced a manual process, ensuring that manufacturers received performance feedback and defect alerts instantly.
• Internal KPI Dashboards: Updated and automated the reporting of internal Quality KPIs sent company-wide via Power BI. These dashboards provided real-time visibility into production health and material performance.
• Project Tracking: Created specialized dashboards for individual product projects, allowing key stakeholders to monitor production progress, identify quality-related delays, and address issues before they reached the warehouse.

Crisis Management: The Skywalker Pro Recall

The role involved a central contribution to one of the company’s most significant safety initiatives: the global recall of the Skywalker Pro Via Ferrata Set.

The Investigation

During routine safety monitoring and stress testing, a potential malfunction was identified in the carabiner’s automatic locking mechanism.

• Defect Confirmation: Instrumental in the laboratory testing phase that confirmed the issue. Through rigorous environmental testing, it was proven that the locking function could fail in cooler temperatures, posing a serious fall hazard.
• Global Task Force: Served as a key technical member of the international task force formed to resolve the crisis. This involved a massive team effort across engineering, legal, and logistics departments in multiple countries.
• Data-Driven Response: Used the newly migrated database to help the task force identify affected batches (specifically 07/2021 and 11/2022). This ensured the recall was targeted and effective, protecting Mammut’s users while maintaining brand trust.

Results

• Successful ERP Transition: Ensured a seamless data migration to Dynamics 365, keeping all quality-dependent programs functional during a critical infrastructure shift.
• Global Safety Impact: The confirmation of the Skywalker defect led to a proactive global recall, preventing potential accidents through data-backed laboratory evidence.
• Efficiency Gains: Automated reporting saved significant weekly hours for the Quality team, allowing them to focus on high-level analysis rather than manual data entry.