Aniket J.

Implemented a production-grade Retrieval-Augmented Generation (RAG) system for intelligent querying over a dataset of 12.4M+ enterprise legal documents. Engineered a scalable pipeline for document ingestion and preprocessing using FastAPI and OCR while supporting semantic chunking and metadata enrichment. Designed a hybrid search architecture combining BM25, dense vector embeddings, and cross-encoder reranking to elevate contextual response accuracy. Developed a robust FastAPI backend with multiple endpoints to manage LLM inference, document retrieval, and user queries, focusing on optimizing response times. Deployed the solution on AWS EKS with Kubernetes autoscaling for high availability and support for concurrent users. Automated infrastructure provisioning using Terraform, which enabled reproducible environments and significant reductions in setup time. Orchestrated the MLOps lifecycle, encompassing Docker containerization, versioning, CI/CD pipelines, and monitoring.

Contributed to the development of a multilingual deepfake speech detection system as part of the Kaggle IndicTTS Challenge. Assisted in creating a scalable audio preprocessing and feature engineering pipeline capable of processing over 33,000 audio clips. Implemented normalization, trimming, channel balancing, and log-scaled Mel spectrogram generation for use in deep learning models. Utilized Vision Transformers (ViT) for classification tasks and participated in training processes using mixed-precision training with multi-GPU setups, enhancing efficiency. Achieved notable improvements in training time while optimizing model performance significantly.