My role involves running distributed compute for ML models. Delivering bug fixes, new features, code reviews, system design and team support.

• LLM Based Retrieval-Augmented Generation with Privacy Monitoring: A RAG system with a Privacy Monitor on AWS for an insurance firm, ensuring data confidentiality and compliance with regulations. Implemented Named Entity Recognition (NER) for secure data redaction and enhanced customer support responses using LLAMA-2, fine-tuned on policy data. ChromaDB was used as a vector store for the storage of embeddings.• Interbank Data Sharing Fraud Detection: Led fraud detection project for Banque de France. Implemented CI/CD pipeline for robust testing. Utilized knowledge graphs, PyTorch for deep learning, Docker for containerization, and GCP for scalability. Achieved 88% F1 score, 20% fewer false positives vs baseline. Demonstrated efficacy of the approach.• Privacy-Enhanced Loss Function for Split Learning: Engineered a novel privacy-centric loss function for split learning models, leveraging the Hilbert Schmidt Independence Criterion a modification to the current best loss function distance decorelation. Enhanced data security by incorporating MixUp, Intra-Instahide, and GradPrune optimization slashing attack success rates by 70% and boosting model convergence efficiency by 33%. The attacker models used are GANs.• Prompt Reconstruction Attack Analysis on Clinical Data: Conducted a successful prompt reconstruction attack on a LLM model in a black-box, split learning environment, utilizing public data. Employed analysis metrics such as BLEU, METEOR, and ROUGE, achieving a 95% success rate in extracting sensitive information from textual data.• Clinical Named Entity Recognition (NER) on Private Medical datasets: Fine-tuned Clinical BERT for NER, initially detecting clinical entities. Enhanced via multi-task joint learning to recognize non-clinical factors such as age, gender, race, and social history. Enables comprehensive understanding of patient health. Configurable, reusable, and scalable for efficient training and inference while ensuring patient privacy.

• Led an NSF-supported collaboration with Solea Energy on a real-time price forecasting system, integrating techniques like Random Forest, Gradient Boosting, and Neural Networks (TCNN, CNN2D, transformers), leading to an 8.33% boost in profit margins across 35 nodes.• Implemented distributed learning with TensorFlow on a lambda cluster, resulting in a 1000% improvement in model training, evaluation, and hyperparameter tuning efficiency.• Developed a novel hybrid feature selection method combining Principal Component Analysis (PCA) and Correlation feature selection. This approach led to a 12% reduction in model error rate.• Effectively communicated complex technical concepts, such as spatiotemporal patterns and congestion pricing factors, to sponsors. Used visual aids and simplified language for clarity, facilitating understanding across diverse expertise levels.• Contributed to and published research findings at the NAPS conference in 2022.