Gustavo Souto

• Building Review Radar: an end-to-end agentic pipeline scraping Google Maps reviews, extracting negative themes via sentiment analysis, and generating operational recommendations. Stack: LangGraph,LangSmith, LLaMA 3.3 70B, FastAPI, Streamlit, Crawl4AI. • Developed a Rossmann Sales Forecasting benchmark exploring time-series modeling and feature engineering for multi-store demand prediction. • Exploring production patterns for agentic systems: stateful graphs, conditional branching, and LLM observability.

• Architected a multi-agent system (LangGraph, A2A protocol) to autonomously crawl and extract structured business intelligence from company websites, delivered directly in client-facing reports. • Designed LLM-based orchestration pipelines (LangChain) to automate data extraction workflows across concurrent projects. • Built and productionized a spatial propensity model integrating geospatial features into a real-time scoring pipeline.

• Deployed time-series forecasting and anomaly detection for 10,000+ machines, processing millions of telemetry records monthly — reducing downtime and cutting false positives by 12%. • Built end-to-end ML pipelines (LightGBM, Isolation Forest, MLflow), accelerating model iteration by 15%. Models in production for 3+ years. • Operational dashboards (Plotly, Streamlit) adopted by Field Application Engineers for fleet decision-making.

• Developed energy disaggregation models (HMMs) to estimate appliance-level consumption from smart meter data, enabling end-users to identify high-consumption devices and prioritize equipment upgrades. • Processed large-scale residential energy datasets using PySpark and Databricks, building scalable pipelines handling millions of household records in production. • Deployed forecasting models via Azure ML, evaluated with MAE and RMSE, optimized for business interpretability for non-technical stakeholders.

• Built churn prediction models across insurance product lines (life, home, auto), increasing customer retention by 32%. • Developed a recommendation engine for cross-selling policies, driving a 14% increase in sales. Designed customer segmentation pipelines (K-Means) to improve targeting efficiency for marketing campaigns. • Engineered scalable ETL pipelines on a Hadoop-based data lake (Hive, MapReduce, Spark) for the full Data Science team.

• Designed hands-on robotics programs for kids and teens that encouraged programming skills, teamwork, and interdisciplinary learning.

• Conducted applied research on real-time traffic anomaly detection using heterogeneous urban data streams. Combining SCATS loop sensors, bus GPS data, and Twitter feeds to identify and diagnose anomalies in Dublin's road network. Built SVM-based classification models on time-series sensor data, contributing to 3+ peer-reviewed publications on event detection and stream processing for smart city applications.

• Lectured undergraduate technical courses in Software Engineering, Web Development, Databases, and Introduction to Programming, covering both theoretical foundations and hands-on lab work. • Mentored students through final technical projects, guiding problem scoping, methodology, and delivery from conception to presentation.

Completed 3 years of doctoral coursework and research in Computer Science, focusing on anomaly detection and event identification in spatial time-series data. Published 4+ peer-reviewed papers on heterogeneous stream processing and urban traffic anomaly detection. Did not complete dissertation.

Research on computer vision systems for real-time remote monitoring of respiratory parameters in ALS patients, detecting risk situations and triggering alerts for healthcare professionals. An early application of sensor-based monitoring and automated anomaly alerting, themes that run through my subsequent work in industrial IoT and predictive maintenance