Vasco – PyTorch, Python, OpenAI

An internationally leading technology enterprise operating in the optics and optoelectronics industries. Vasco has worked on three projects:

• Time Series Forecasting (Python, Generative AI, Time Series)
• Manufacturing Process Optimization (Python, Causal Discovery, Causal Inference)
• Log Analytics Anomaly Detection (Python, Generative AI, Time Series)

• Led time series forecasting initiatives across diverse departments, acting as the lead technical stakeholder and driving the adoption of advanced forecasting methodologies, including Generative AI-based approaches to enhance prediction accuracy and scenario modeling for financial forecasting.
• Led and mentored a team of data scientists and machine learning engineers through the full lifecycle of machine learning products.
• Functioned as a solutions architect for inter-departmental ETL and data architecture initiatives.
• Led application of advanced causal discovery (PC/FCI, GES, LiNGAM) on high-dimensional sensor data, uncovering latent defect structures.
• Architected and implemented doubly robust machine learning solutions (DML, DR-Learners) for unbiased causal eOect estimation of manufacturing parameters on lens scrape rate (MTF, defects).
• Developed and evangelized the use of Bayesian Networks derived from causal DAGs for 'what-if' intervention simulations (Do-calculus), enabling proactive quality control strategies adopted by product teams.
• Pioneered and scaled a reusable causal inference pipeline (integrating DoWhy/EconML) for automated ATE/CATE estimation.
• Architected and led the development of an enterprise-scale log analytics platform, leveraging advanced unsupervised learning (e.g., autoencoders, isolation forests, clustering) for real-time, multi-source anomaly detection, significantly reducing Mean Time to Detection (MTTD) by 4%.
• Led the integration of NLP techniques (e.g., log parsing, embeddings, topic modeling) with deep learning models (e.g., LSTMs, Transformers) to uncover nuanced anomalies in unstructured log data.
• Engineered a modular, A/B testable framework for deploying and evaluating diverse anomaly detection algorithms (statistical, ML-based) on high-velocity streaming logs, enabling proactive alerting and automated incident correlation which reduced alert fatigue by 7%.
• Orchestrated the design and deployment of a robust MLOps pipeline.

A digital advertising platform focused on optimizing campaign performance and maximizing ROI. It leveraged advanced algorithms and data-driven insights to deliver targeted ads across multiple channels. It helped businesses achieve greater reach, engagement, and efficiency in their advertising efforts.

Vasco was responsible for the following:

• ideated and developed AI-driven product solutions tailored to client needs, leveraging LangChain and LlamaIndex for LLM application development and knowledge integration;
• designed and deployed Retrieval Augmented Generation (RAG) pipelines to improve information retrieval and contextual responses, using Pinecone and Weaviate for vector search optimization;
• built and fine-tuned Large Language Models (LLMs) for chatbot applications, utilizing OpenAI APIs, prompt engineering, and generative AI frameworks;
• created scalable RESTful APIs with FastAPI to deploy AI and ML services, ensuring seamless integration with existing systems;
• deployed generative AI models in production, focusing on quantization, inference optimization, and deployment using PyTorch, TensorFlow, and Hugging Face Transformers;
• applied computer vision techniques for creative tagging and performance analysis, leveraging CLIP, Deep Learning, and Keras, driving campaign ROI improvements;
• designed and implemented MLOps workflows, including monitoring, versioning, and scaling AI systems with Kubernetes, Docker, and cloud platforms like AWS EKS, GCP, and Azure;
• developed robust data pipelines for real-time analytics and machine learning workflows, integrating BigQuery, DBT, and Databricks to enhance performance and scalability;
• integrated vector databases (e.g., Pinecone, Weaviate) and embeddings for advanced semantic search and retrieval in AI-driven applications.

A solution that automates the generation of AdWords text and optimizes bidding strategies for recruitment campaigns. Designed to enhance performance, it streamlines ad creation while ensuring cost-effective delivery to the right audience. This tool helps employers maximize their recruitment ROI with minimal effort.

Vasco's main responsibilities included:

• developed and architected very complex real-time ML systems multiple times to production as a team;
• developed and managed an automated bidding model optimized to maximize revenue/margin, generating a 6% uplift in spend and a decrease in CPA;
• developed and managed an automated text generation models (keywords, ad copy) using classic NLP and LLM techniques (self-hosting and external, prompt engineering, persuasion techniques);
• addressed third-party cookie deprecation by using probabilistic clustering in full-funnel attribution, using Spark to pipeline data to enhance conversion tracking and data accuracy;
• leveraged causal inference (double robust learners — econml) to refine conversion tracking, boosting accuracy and stakeholder confidence on top of clustering output;
• developed, A/B tested, and deployed with Docker/Kubernetes a linear regression-based algorithm, achieving a 3.7% YoY revenue increase and enhancing pacing KPIs by 35%;
• successfully led A/B testing for new product features, driving improvement in core company KPIs;
• collaborated with cross-functional teams to enable data-driven decision-making and stakeholder buy-in; recruited elements for the DS team.

A solution that automated the generation of AdWords text and optimized bidding strategies for recruitment campaigns. Designed to enhance performance, it streamlined ad creation while ensuring cost-effective delivery to the right audience. This tool helped employers maximize their recruitment ROI with minimal effort.

Vasco carried out the following responsibilities:

• created scalable RESTful APIs with FastAPI to deploy AI and ML services, ensuring seamless integration with existing systems;
• deployed generative AI models in production, focusing on quantization, inference optimization, and deployment using PyTorch, TensorFlow, and Hugging Face Transformers;
• applied computer vision techniques for creative tagging and performance analysis, leveraging CLIP, Deep Learning, and Keras, driving campaign ROI improvements;
• designed and implemented MLOps workflows, including monitoring, versioning, and scaling AI systems with Kubernetes, Docker, and cloud platforms like AWS EKS, GCP, and Azure;
• developed robust data pipelines for real-time analytics and machine learning workflows, integrating BigQuery, DBT, and Databricks to enhance performance and scalability;
• integrated vector databases (e.g., Pinecone, Weaviate) and embeddings for advanced semantic search and retrieval in AI- driven applications.

A technology-driven platform offered advanced solutions for data analysis, visualization, and decision-making. It combined cutting-edge tools with intuitive interfaces to empower businesses to transform complex data into actionable insights. The platform supported a wide range of industries, enabling smarter strategies and enhanced outcomes.

Among Vasco's contributions, there were the following:

• developed a system to optimize portfolio risk: key risk KPIs (implied volatility, maximum drawdown) decreased on average 7.6%;
• delivered a system for option pricing using deep learning, leading to an improvement in average daily returns of 2.7%;
• managed end-to-end data processing, enhancing system performance through effective sourcing, preprocessing, and partitioning for model training and inference;
• utilized NLP techniques to analyze tweets, creating a feature store (embedding) for machine learning models that enhanced stock market understanding;
• developed a back-and-forward testing framework.