• Building Evals for search & browse Agents.

• Working on Video Search and Ranking/relevance at Prime Video

• Designed contrastive reward modeling and PPO-based alignment pipeline (RLSF) to ground responses of reasoning model in corpus human preference data, adapting RLHF for synthetic feedback alignment.

• Integrated Fully-Sharded Data Parallelism(FSDP) for Supervised Fine-Tuning (SFT) and Reinforcement Learning with Synthetic Feedback (RLSF) to scale Post-training of 70B Model over distributed AI infra.

• Orchestrated the scalable model serving of 405B LLM over multi-node Kubernetes clusters using Ray and vLLM, optimizing inference latency to handle high-volume of requests in production.

• Enhanced domain-specific answering capabilities with reasoning to improve the knowledge depth of small language models for context-aware question-answering.

• Engineered multi-phase CoT-based data generation pipeline for domain-specific Retrieval Augmented Fine-Tuning (RAFT), resulting in a 12% cost reduction in the synthetic data generation pipeline.

• Proposed multi-conditioned iterative latent variable refinement-based DDPM for image-to-image for CT Image generation.

• Engineered novel training-free style injection method to significantly improve image-guided latent refinement in mammogram patch synthesis.

• Dedicated to refining the noise scheduler for sampling procedures and mitigating signal-to-noise ratio to (0.12%) in the forward diffusion process.

• Collaborated with the team of 4 mlops engineers and delivered end-to-end ML monitoring platform, including building feature engineering pipelines and performing inference of adversarial detection.

• Built touchless ML model productionalization pipeline for real-time ML inference services, enhancing streamline model lifecycle management, versiong and experiment tracking, which improved deployment speed by 15% and optimised ML workflows in production environment.

• Developed 5% enhanced LoRA-based contextual embedding model to improve drift detection algorithm to enable precise detection of data-distribution shifts in data embeddings.

• Rolled updated version of feature store database and scalable data ingestion infrastructure using Apache-Kafka and pySpark, resulting in a 8% boost in high-performance data processing.

◦ Researched various Drifts and their detection algorithms.

◦ Developed end to end Adversarial and Drift detection mechanism and deployed it into the production pipeline using MLOps principles