Giovanni Belval

• Built an AI agent monitoring warehouse operations, improving fulfillment responsiveness by 17%.
• Designed an LLM-as-a-judge evaluation pipeline to benchmark agent reliability and reduce hallucinations.
• Fine-tuned a YOLO model for barcode detection, cutting manual lookup time by 32%.

• Built a data synthesis pipeline for generating realistic documents using fine-tuned LLMs while preserving privacy.
• Implemented memory-efficient inference and deployed scalable Vertex AI endpoints on GCP.

• Developed a "super learner" stacking ensemble model (93% accuracy) for vehicle total loss prediction.
• Packaged the algorithm into a reusable internal AutoML library.

Investigation of Multi agent debate performances.

Generated key robot motion sequences using an open-source video-generation model, enabling their downstream use as policy inputs for robotic control.

Worked on a novel hierarchical Hidden Markov Model architecture that is highly parallelizable for high speed and accuracy, significantly outperforming standard Markov chain approaches.