Paulo Soares

Pinterest

• ▹Optimized large embedding table model through feature expansion and ablation studies; launched to production with −1.34% CPC, +2.6% CTR, and six-figures annual infra savings.
• ▹Built end-to-end pre-trained entity embedding system adopted into production Home Feed model, improving engagement by +4.6% and session success by +0.24% with neutral latency.
• ▹Integrated cross-domain embeddings into ads ranking model, driving −0.54% CPC, +1.32% CTR, and seven-figures annual infra savings through feature optimization.
• ▹Created ML experimentation guidelines and delivered technical deep dives, establishing best practices for model development and production maintenance.

Tetricus Labs

• ▹Prototyped synthetic conversational data generation using LLM-empowered agents and fine-tuned models for realistic, stylistically tailored outputs; drafted scientific manuscript for publication.

Pinterest

• ▹Developed end-to-end model distillation and domain adaptation solutions for large-scale ad ranking systems, applying statistical and deep learning techniques to reduce bias.
• ▹Designed and implemented big data pipelines for daily model training, inference, and reporting using distributed computing frameworks.
• ▹Created technical documentation, design specs, and onboarding tutorials that accelerated team development velocity and supported knowledge transfer.

Petrobras

• ▹Led $5M project to expedite maintenance and drilling operations planning, delivering expected $15M annual savings.
• ▹Engineered ship usage optimization system, increasing operational efficiency by 15% through reduced downtime.