Generative AI & Machine Learning Projects Portfolio

2025 2

1. AI for Legal (In Production) 2

2. Email Classifier for Back Office (In Production) 3

3. Injury Prevention & Performance Optimization Platform for a Professional Football Club (MVP in Production) 4

4. Professional Football Club – Generative AI for Marketing Content (Experimental, Negative Example of PoC Results) 5

5.  Customer FAQ Chatbot (POC) 7

6. Ticket Categorization and Knowledge Base Foundation (Poc Succeed, MVP testing) 8

7. Service Desk Assistant with Jira Integration (POC) 9

2024 10

1. Global Product Recommendation Engine for a Luxury Fashion Brand (Relevant, In Production) 10

2. International Luxury Brand – Generative AI for Multilingual Marketing (In Production) 11

3. Luxury Jewelry Brand – Generative AI for Design and Cataloging (Experimental, Negative Example of PoC Results) 12

4. Industrial Manufacturer – AI for Commercial Offer Ranking (GenAI, Good example in manufacturing) 13

5. Healthcare Tech Company – Spillage Detection (MVP Succeeded) 14

6. Media Platform – Talent Matching & Content Moderation (MVP Succeeded) 15

2025

1. AI for Legal (In Production)

Situation:

The legal team at a major organization faced a significant challenge in keeping up with weekly updates from official sources like the Gazzetta Ufficiale and Eur-Lex. Regulatory documents were lengthy (200+ pages per week) and constantly evolving due to new EU regulations, such as GDPR updates and other compliance mandates. Reviewing these manually required 6–8 hours per document, with coverage often limited to 65%, increasing the risk of missing critical changes and impacting business compliance. Additionally, the team needed a way to stay proactive in monitoring legal updates across multiple jurisdictions while integrating feedback into the workflow.

Task:

Design an end-to-end AI-powered solution that:

• Automatically extracts, summarizes, and prioritizes new legal content.

• Supports interactive exploration and comparison of legal documents.

• Ensures alignment with the latest regulations and GDPR requirements.

• Keeps the legal team actively involved in the approval process to maintain trust and oversight.

• Reduces manual workload, improves coverage, and provides actionable guidance on relevant regulatory updates.

Action:

• Developed a multi-agent architecture:

  • Interactive Chatbot Agent: Allows users to upload documents, query content, perform semantic searches, and compare regulations side-by-side.

  • Task-Specific Experimental Agent: Designed to adapt AI behavior for specialized legal tasks, such as extracting GDPR-specific clauses or flagging changes in EU directives. (Experimental)

• Built an MVP interface with guided workflows for manual approval, ensuring legal team members were involved at every step of AI-generated summaries and recommendations (Replit, React-Based)

• Implemented continuous monitoring and update pipelines to automatically incorporate changes from official sources and keep the system aligned with evolving regulations (e.g., GDPR updates).

• Orchestrated backend using Google Cloud Functions,Workflows, BigQuery, and Cloud Storage, while integrating LLMs (Gemini, GPT) for relevance scoring, summarization, and semantic embeddings.

• Enabled feedback loops for continuous model improvement, with user feedback directly impacting the AI’s relevance scoring and summarization accuracy.

Result:

• Reduced document review time by ~70% (from 6–8 hours to 1–2 hours per document).

• Increased coverage of incoming regulatory documents from 65% to over 90%.

• Annual savings in manual effort estimated at €150k–€180k.

• 80% of the legal team reported improved efficiency; 65% noted clearer actionable guidance.

• The system is fully compliant with GDPR and new EU regulations, providing confidence that updates are integrated in real-time and that the legal team remains in control.

• Enabled a proactive approach to regulation monitoring, ensuring the team always has the latest summaries and insights ready for decision-making.

Technologies: Google Cloud (Cloud Functions, BigQuery, Cloud Storage), React-based web application (Replit), Gemini, GPT, multi-agent orchestration.

2. Email Classifier for Back Office (In Production)

Situation:

The back-office team managed multiple shared email inboxes from over 40–50 branches nationwide. Emails included reimbursements, bonuses, school expenses, contracts, promotions, and general inquiries. The legacy system was rule-based, error-prone, and required high manual effort, leading to slow response times and inconsistent categorization.

Task:

Develop an intelligent email routing system that:

• Automatically classifies incoming messages with high accuracy.

• Reduces manual workload while maintaining traceability.

• Handles complex or ambiguous cases using AI.

Action:

• Designed a hybrid architecture combining:

  • Customizable rule engine for straightforward email routing.

  • Generative AI classifier (Gemini) for ambiguous or multi-topic emails.

• Implemented automated validation and error logging to track misclassifications and allow continuous improvement.

• Integrated the system across all back-office email flows with secure access and compliance protocols.

Result:

• Achieved 99.57% classification accuracy across 1,247+ emails (target was 95%).

• Reduced manual email handling time by ~80%.

• Estimated 70% cost reduction versus legacy API solutions.

• Improved traceability and accountability in back-office operations.

Technologies: Gemini, Python, custom rules engine, backend orchestration, monitoring scripts.

3. Injury Prevention & Performance Optimization Platform for a Professional Football Club (MVP in Production)

Situation:

A top-tier European professional football club (NDA-protected name, annual revenue ~€500M) aimed to maximize player performance while minimizing injury risk. Players wore wearable devices capturing hundreds of physiological parameters during training and matches. Manual data monitoring and integration with medical reports was slow and error-prone.

Task:

Build an AI-powered platform to predict injury risk, recommend training adjustments, and provide actionable insights, integrating wearables, biomedical data, and historical medical records. Include a generative AI interface for querying player status, recovery progress, and risk explanations.

Action:

• Developed predictive models using Vertex AI and AutoML with multimodal data.

• Built an Injury Risk Engine to compute player-specific risk scores and training recommendations.

• Integrated a Generative AI web app for exploratory analysis, natural language queries, and simulation of training adjustments.

• Implemented data pipelines for wearable telemetry and historical medical documents using semantic search and embedding-based retrieval.

• Enabled feedback-driven fine-tuning of AI suggestions via medical and coaching staff review.

• Created visual dashboards for trends, risk correlations, and recovery projections.

Result / Impact:

• 25% reduction in preventable injuries.

• ~40–50% time savings for medical staff, measured as reduction in hours spent reviewing and integrating player data per week (from ~3–4 hours/player to ~1–1.5 hours).

• ~45% of AI-generated recommendations approved without full manual review, reducing workload and allowing staff to focus on high-priority decisions.

• Real-time insights and generative AI queries enabled better understanding of injury causes, monitoring of recovery, and adaptive training plans.

• Scalable platform capable of incorporating new biomedical data and adjusting training strategies accordingly.

Technologies:

Vertex AI, AutoML, Generative AI (Gemini), Cloud Storage, React dashboards (Replit), semantic search & embeddings, continuous feedback integration.

4. Professional Football Club – Generative AI for Marketing Content (Experimental, Negative Example of PoC Results)

Situation:

The club, aiming to maximize player performance and health, also maintains a large marketing team responsible for content across social media, fan engagement, and campaign materials. There was growing interest in applying generative AI to automate and accelerate content creation. At the time, generative models (Imagen 2, Imagen 3, VEO 1 – Preview) were rapidly evolving, and there was significant hype around AI-driven creative workflows.

Task:

• Experimentally support the marketing team with AI-assisted content generation for campaigns, social posts, and fan engagement materials.

• Test AI capabilities while assessing limitations and integration challenges with the existing web-based content management platform.

Action:

• Integrated experimental generative AI pipelines using Imagen 2, Imagen 3, and VEO 1 for image and text content creation.

• Developed an internal interface allowing marketing staff to input prompts and receive AI-generated drafts.

• Monitored output quality and prompt sensitivity, and documented failure modes for continuous learning.

• What we could have done better : Conducted informal training sessions and workshops to help the team understand model behavior, limitations, and best practices, since initial expectations were that they would test the system independently.

• What we could have done better : Implemented tracking for content reuse, editing effort, and feedback loops, highlighting where AI outputs could be leveraged or required manual refinement.

Result:

• Partial successes: Some generated content (e.g., simple banners, fan graphics) was usable after minimal editing.

• Limitations observed:

  • Inconsistent quality, especially for complex prompts or brand-specific design requirements.

  • High latency and occasional service interruptions.

  • Misalignment with marketing tone and style, requiring human oversight.

• Learning outcomes:

  • Highlighted the gap between hype and technology maturity, emphasizing the need for staff training before full deployment.

  • Paved the way for future upsell/resell strategies using more mature models (e.g., Nano Banana), leveraging the platform infrastructure already built.

  • Reinforced the importance of evangelization and change management in adopting generative AI tools.

Technologies:

Generative AI models (Imagen 2, Imagen 3, VEO 1 – Preview), internal web-based content management interface, workflow monitoring, feedback tracking.

Key Takeaways:

• Even with advanced AI, human oversight and training are critical for practical deployment.

• Early experimentation informs future adoption strategies, particularly for fast-evolving generative models.

• Managing expectations around AI hype is essential to avoid premature investment and frustration.

5.  Customer FAQ Chatbot (POC)

Situation:

The e-commerce platform’s FAQ system was static and insufficient, resulting in a high volume of level-1 customer service tickets. Customers struggled to find relevant answers quickly, reducing self-service efficiency and satisfaction.

Task:

Create an AI-powered chatbot that:

• Provides contextual, real-time responses based on historical tickets.

• Identifies gaps in the FAQ knowledge base and suggests expansions.

• Reduces service ticket volume and improves self-service experience.

Action:

• Developed a ticket analysis engine to extract recurring questions and categorize them.

• Built FAQ enrichment logic to recommend new content and improve coverage.

• Integrated natural language generation to deliver user-friendly, contextual answers.

• Designed frontend chatbot interface for web and mobile platforms.

Result:

• Reduced level-1 customer service tickets by 40–50%.

• Achieved ~30% increase in self-service completion rate.

• Monthly enrichment cycles improved FAQ coverage and relevance.

• NPS improved by 10–15 points.

Technologies: Gemini, NLP, Python, React-based chatbot interface, historical ticket dataset.

6. Ticket Categorization and Knowledge Base Foundation (Poc Succeed, MVP testing)

Situation:

Customer service tickets were scattered across multiple platforms (Salesforce, Jira, email), with inconsistent categorization and metadata. This hindered automation, analytics, and chatbot integration.

Task:

Create a unified foundation for ticket classification and knowledge management:

• Standardize ticket taxonomy across channels.

• Cluster recurring issues and align metadata.

• Prepare a “golden dataset” for future AI-based automation.

Action:

• Built cross-platform classification and clustering models.

• Aligned ticket ontology across departments.

• Extracted golden dataset from historical email flows for training future models.

Result:

• Achieved 90%+ coverage of top ticket categories.

• Reached 85%+ initial classification accuracy.

• Enabled automation for ~35% of recurring ticket flows.

• Improved visibility on volume, trends, and urgency by 70%.

Technologies: Python, clustering algorithms, embeddings, LLMs for classification, data pipelines.

7. Service Desk Assistant with Jira Integration (POC)

Situation:

Internal service desk agents faced repetitive first-line tasks, slowing response times and onboarding of junior agents. Manual management of Jira tickets was prone to delays.

Task:

Create AI assistants to:

• Automate first-line support for internal users.

• Enable semi-automated Jira ticket actions.

• Detect trending issues in real time.

Action:

• Developed custom Jira connector for cloud and on-prem.

• Integrated knowledge base retrieval system for agent support.

• Built semi-automated action engine for ticket escalation, closure, and classification.

Result:

• Automated 45–60% of first-line responses. ( raw estimation )

• Reduced average ticket resolution time by ~35%. ( raw estimation )

• Real-time detection of trending issues improved operational awareness.

Technologies: Jira APIs, Python, LLMs, semantic search, React-based agent interface.

2024

1. Global Product Recommendation Engine for a Luxury Fashion Brand (Relevant, In Production)

Situation:

A leading international luxury fashion brand ( Annual revenue ~€10B) relied on a legacy, third-party recommendation system with high licensing costs (~€0.5M/year) and limited flexibility. The system was not scalable globally, lacked integration with in-store operations, and could not adapt quickly to user behavioral changes or new product lines. The brand aimed to increase personalized engagement, conversion rates, and units per transaction across e-commerce and physical stores.

Task:

Design and implement an internal recommendation engine capable of:

• Supporting sales assistants in the USA for real-time, personalized suggestions.

• Extending to e-commerce platforms across Europe and eventually Asia.

• Providing actionable insights and generative recommendations for new collections.

• Ensuring flexibility, scalability, and data transparency while reducing dependency on expensive third-party solutions.

Action:

• Built a machine learning & deep learning recommendation system using Databricks for data processing and AWS backend for cloud infrastructure.

• Implemented streaming web applications for sales assistants, allowing real-time interaction and feedback on recommendations.

• Deployed A/B testing frameworks to measure uplift in conversion rate, units per transaction, and customer engagement.

• Introduced automated data ingestion flows, quality checks, and monitoring, enabling the platform to support 30+ data scientists working in parallel.

• Designed generative AI modules to summarize customer preferences, propose complementary products, and facilitate personalized recommendations across multiple channels.

• Coordinated phased rollout: first tested internally with sales assistants, then scaled to global e-commerce operations, optimizing for multi-region deployment.

Result / Impact:

• ~12% increase in conversion rate across tested markets (USA & Europe).

• Uplift in units per transaction (UPT) ~10%, with clear statistical significance.

• Reduction in licensing costs by ~€0.5M/year, replaced by a fully owned system.

• Scalable platform currently deployed across the USA and parts of Europe, with Asia expansion in progress.

• Platform supports real-time recommendation delivery, improved operational efficiency, and actionable insights for marketing and sales teams.

• Generative AI integration enables sales assistants and online users to access tailored suggestions and enhanced product discovery.

Technologies:

Databricks, AWS (compute, storage, monitoring), ML & deep learning frameworks (PyTorch / TensorFlow), streaming web applications, automated data pipelines, generative AI modules.

2. International Luxury Brand – Generative AI for Multilingual Marketing (In Production)

Situation:

A global luxury brand (NDA-protected) ran marketing campaigns across multiple regions, requiring fast, accurate, and culturally adapted multilingual translations. The campaigns involved complex terminology, precise brand vocabulary, and rules that varied by market—for example, some terms had to remain in English or Italian, while others followed strict conventions in French or Asian markets.

Task:

Automate translation, adaptation, and quality control of marketing content for global campaigns, ensuring consistency, accuracy, and compliance with brand-specific terminology.

Action:

• Implemented Vertex AI generative models for translations and contextual content adaptation.

• Built a modular AI engine capable of handling market-specific rules and exceptions.

• Integrated the solution directly into campaign pipelines, enabling real-time translation checks and adjustments for regional campaigns.

• Conducted iterative validation with marketing and localization teams to ensure accuracy and alignment with brand standards.

Result / Impact:

• Accelerated campaign rollout by ~50% compared to manual processes.

• Ensured high translation accuracy and consistency across all regions.

• Reduced manual localization effort by ~60%, freeing resources for creative tasks.

• Successfully handled complex multilingual requirements across Europe, Asia, and North America.

Technologies:

Vertex AI, Generative AI models, pipeline automation, market-specific rules engine.

3. Luxury Jewelry Brand – Generative AI for Design and Cataloging (Experimental, Negative Example of PoC Results)

Situation:

The client, a high-end jewelry brand, had 20 years of design sketches, which were manually stored and extremely difficult to navigate. Inspired by the hype around generative AI, they were interested in exploring AI-assisted design generation, but the technology was still in early preview stages.

Task:

1. Cataloging: Automate classification and organization of historical sketches for easier retrieval.

2. Generative Design (POC): Experiment with AI to generate new design sketches that could be potentially used in collections.

Action:

• Cataloging:

  • Built classification models using Vertex AI and AutoML.

  • Small-scale labeling by business users enabled AI to categorize sketches by type, collection, and design features.

  • Created a searchable interface for designers to explore historical sketches.

• Generative Design (POC):

  • Leveraged early-stage generative models (e.g., Imagen 2, Oveo – Preview).

  • Implemented experimental pipelines for sketch generation and refinement.

  • Integrated feedback from design teams to evaluate quality and feasibility.

Result:

• Cataloging:

  • Successfully indexed ~20,000 sketches from 20 years, reducing manual search time from weeks to hours.

  • Designers could retrieve sketches ~80% faster, improving workflow and enabling data-driven inspiration.

• Generative Design (POC):

  • AI-generated sketches did not meet brand quality standards due to model immaturity and lack of fine-grained control over design details.

  • Despite partial successes, the technology was not ready for production, but the POC provided critical insights into limitations and areas for future improvement.

Technologies:

Vertex AI, AutoML, Generative AI (Imagen 2, Oveo – Preview), AI-assisted sketch classification and experimentation.

4. Industrial Manufacturer – AI for Commercial Offer Ranking (GenAI, Good example in manufacturing)

Situation:

The industrial manufacturer manages over 10,000 commercial offers annually, with a large sales team operating autonomously. Each sales representative handled clients individually, often prioritizing relationships over data-driven decisions. The board had no visibility into sales performance, offer management, or efficiency, making it extremely difficult to optimize engagement and revenue. Manual offer ranking and preparation were slow, inconsistent, and prone to missed opportunities, especially for lower-value projects.

Task:

• Provide a centralized, transparent system for both the board and sales teams.

• Prioritize offers based on predicted likelihood of winning and potential revenue impact.

• Automate the generation of offer documents using generative AI, leveraging historical offers and client-specific data to reduce preparation time.

• Ultimately, increase sales efficiency and improve revenue capture from previously overlooked opportunities.

Action:

• Developed a web-based portal with two access levels: one for the board to monitor sales activity and performance metrics, and one for sales representatives to manage their own client pipeline.

• Implemented a machine learning engine to predict the probability of winning each commercial offer based on historical data, client behavior, and project characteristics.

• Integrated generative AI to automatically draft offer documents, using prior offers from the same client or similar clients as reference.

• Conducted training sessions for sales teams to ensure smooth adoption and maximize engagement.

• Built tracking and analytics dashboards to monitor offer submission rates, predicted win probabilities, and actual outcomes, providing actionable insights for both sales and management.

Result:

• 50–60% reduction in manual time spent preparing and sending offers.

• 25% increase in offers submitted for low-value or previously neglected projects.

• Improved win rate prediction accuracy to ~78%, allowing sales teams to focus efforts on high-impact opportunities.

• Generated an estimated 10–15% uplift in annual revenue by capturing more opportunities that would have been previously overlooked.

• The platform proved scalable and was resold to a similar industrial client, demonstrating its robustness and applicability across manufacturing organizations.

Technologies:

Data science models (Python, scikit-learn, XGBoost), generative AI for document creation (Vertex AI / GPT-based), web portal with dashboards, analytics tracking, and automated workflow orchestration.

Key Takeaways:

• Centralizing offer management and combining predictive analytics with generative AI dramatically improves efficiency and revenue capture.

• AI-generated documents reduce human bottlenecks, especially for repetitive or low-value tasks.

• A transparent portal empowers both sales and management, aligning incentives and providing data-driven decision-making.

5. Healthcare Tech Company – Spillage Detection (2024)

Situation:

Lab operators manually monitored continuous video footage from test tube handling to detect spillage. During night shifts, no one was present, so operators had to review tens of hours of video retrospectively, risking safety incidents and delaying experiments.

Task:

• Automatically detect spillage events in real-time or near real-time.

• Reduce manual video review and improve lab safety.

Action:

• Developed an automatic spillage detection pipeline using Google Cloud Vertex AI and Cloud Vision object detection.

• Created a web app with Streamlit dashboards, allowing operators to review flagged clips efficiently and confirm events.

• Orchestrated the workflow with Cloud Run to handle video ingestion and processing seamlessly.

Result:

• Significantly reduced manual review time, enabling operators to focus only on flagged events.

• Enhanced lab safety and operational efficiency, ensuring timely detection of spillage incidents.

Technologies:

Vertex AI, Cloud Vision, Cloud Run, Streamlit dashboards, Python.

6. Media Platform – Talent Matching & Content Moderation 

Situation:

Connecting clients with actors and managing user-generated content manually was slow, inconsistent, and error-prone, impacting user experience and platform safety.

Task:

• Automate talent matchmaking for clients.

• Ensure content moderation is accurate, scalable, and real-time.

Action:

• Developed machine learning models to match clients with actors based on profile, skills, and preferences.

• Deployed AI content moderation pipelines to automatically flag inappropriate or unsafe content.

• Integrated the solution into the platform workflow, reducing manual intervention.

Result:

• Enhanced user experience with faster, more relevant talent matches.

• Improved platform safety and compliance through automated moderation.

• Freed human moderators to focus on complex cases.

Technologies:

Machine Learning, AI content moderation pipelines, platform integration.

Pre 2024 Projects

Falck Renewables (Pre-2024) – Predictive Maintenance for Wind Turbines

Developed a machine learning system on AWS to predict maintenance needs for a wind farm consisting of 50 wind turbines. The project improved operational efficiency and reduced downtime by predicting potential failures before they occurred.

Technologies: AWS, Predictive Maintenance AI

Automotive Company (Pre-2024) – ML Engineering for Data Science Team

Designed machine learning pipelines to support a team of 30 data scientists, improving their ability to deploy models and manage data workflows efficiently.

Technologies: Machine Learning, Data Pipelines

Postal Services Company (Pre-2024) – Hierarchical Cash Flow Forecasting

Developed a predictive model for hierarchical cash flow forecasting, improving financial planning and resource allocation processes.

Technologies: Predictive Analytics, Data Science Models