Thales Pomari

Infrastructure to deploy, monitor, and manage ML models in production: end-to-end automated pipelines (feature engineering, deployment, and continuous evaluation), CI/CD for reproducible model releases, and monitoring, logging, and alerting — partnering with data scientists and product teams.

• Led a squad of 5 to 8 data engineers: technical and strategic team management, solution architecture, and project management over a ~900 TB data lake (AWS, medallion architecture, Glue Workflows, Jobs, and Data Catalog).
• Led the integration of a new policy sales and claims management system (life insurance with investment product): business rules review, data modeling and mapping, and ETL development into the lake standard.
• Maintained and evolved a CDC pipeline with Apache Kafka for customer, policy, and complaint data, processed and served through DynamoDB with APIs consumed by the customer service CRM.
• Deployed ETL observability from scratch: a Python library for Glue jobs with Lambda-triggered alerts in Microsoft Teams (error details and CloudWatch log link), moving the team from incident-driven detection to proactive resolution before users noticed issues.
• Led the data engineering and MLOps work of the company's first AI project, a decision-support solution for claims analysis: aggregating customer, policy, and complaint data, integrating with the classification model developed by the data science team, and delivering grounded recommendations in a Power BI dashboard.
• Managed the Open Insurance integration: receiving data from other insurers and submitting data in the regulatory standard required by SUSEP.
• Enabled the ML platform for data scientists: SageMaker AI Notebooks with lifecycle scripts integrated with the data lake.
• Created automated deployment from scratch with AWS CodeBuild: backups, artifact generation, automatic change documentation, and publishing to S3, replacing a fully manual process.
• Restructured the team's code governance, previously treated as simple backup: new repositories with CloudFormation templates (infrastructure as code), conventional commits, and a feature/release branching flow, replacing direct commits to dev.

• Built the production pipeline for 3 open-source LLMs (Gemma, Llama, and Qwen) to productionize a document analysis solution.
• Developed prediction serving and management APIs, with model and version orchestration.
• Manage the cloud infrastructure serving models to external customers.
• Integrate generative AI into the engineering workflow (Claude and agents): code review, code quality and security, code standardization, and logging, reducing review time and documentation dependency. Implemented unit and end-to-end tests with Claude assistance.

• Implemented and maintained CDPs (Adobe Experience Platform / Real-Time CDP) for large telecom, retail, and banking companies; the main project served a telecom carrier with roughly 80 million customers.
• Modeled XDM schemas with schema evolution and built batch data ingestion (BigQuery connection and CSV file mapping) and streaming ingestion, configuring client-side Kafka integrated with AEP.
• Used AEP APIs to query customer, campaign, and segment data.
• Developed analytical reports through Query Service (PostgreSQL with nested JSON structures): parsing data to track each customer's campaign touchpoints, reproducing Adobe's native graphical view with more precise filters.
• Created multichannel activation journeys in Adobe Journey Optimizer (email, push, and SMS), including domain IP warming strategies, and contributed to the native Zenvia-Adobe integration for SMS delivery, specifying endpoints and documenting the data exchange between systems.
• Built data pipelines with Airflow and Redshift across AWS and GCP environments.

• Created an internal Feature Store for registering and managing models and variables (SQL and metadata, with edit and delete), with OAuth 2.0 authentication integrated with GCP and a Bootstrap frontend.
• Refactored similarity and segmentation variable calculations from BigQuery to Spark/Scala on Dataproc (10 to 15 node cluster): runtime reduced from days, with failures and restarts, to about 4 hours, cutting cloud costs.
• Developed REST APIs in Flask and Airflow DAGs with Python jobs for data processing and model and variable calculation.
• Deployed through CI/CD and maintained projects on GCP: Kubernetes, Dataflow, Bigtable, BigQuery, and Cloud Storage.

• Developed Java MapReduce routines, running on on-premise servers, to generate prepaid and postpaid customer indicators for the marketing team of a major telecom carrier, processing ETLs of roughly 4.5 billion rows, with per-environment algorithm tuning.
• Supported business rule definitions, mapping the paths between tables to reach the information requested by the marketing team.
• Performed loads and optimizations on Hive and Greenplum through Apache Sqoop, produced analyses and reports, and maintained SAS scripts.