Jio Platforms Ltd | Data Scientist | Enterprise AI Platforms (Aug’23 - Present)

• Real-Time TTS Engine: Delivered cross-lingual TTS APIs with efficient GPU utilization & ultra-fast response times
• Orchestrated distributed training on 4 NVIDIA A100 GPUs using PyTorch DDP, reducing training time by 40%
• Employed CUDA Graphs to eliminate repetitive execution overhead, achieving 0.8s inference for 15-token inputs
• Enabled 50+ concurrent TTS streams via NVIDIA MIG instances, maintaining 0.2 RTF at 90% GPU utilization
• Developed a Python API wrapper to orchestrate LLaMA-driven semantic tokenization & VQGAN-based vocoder
• Deployed Gunicorn/Nginx with round-robin distribution, cutting API errors to 0.5% during peak traffic spikes
• Multilingual OCR System: Advanced Regional Multi-Language OCR with High-Accuracy Language classifier model
• Fine-tuned PaddleClas PULC based language classifier model for 13 languages, achieving a 93.54% F1 score
• Designed multi-stage OCR pipeline with adaptive routing using confidence scores, improving the accuracy by 25%
• Validated the OCR pipeline at enterprise scale, delivering 85% accuracy for 13 languages in production workflows
• Object & Landmark Detection System: Advanced Multi-Object and Landmark Recognition with Scalable Deployment
• Fine-tuned RT-DETR (AP = 0.884 for object detection) & DINOv2 (93.51% F1 score for landmark classification)
• Improved the model performance through grid search optimization of learning rate, momentum, & weight decay
• Designed high-performance pipeline with Flask API, JWT, batch processing, and optimized for real-time response
• Devised scalable Docker/Kubernetes deployments with HPA auto-scaling & Logstash, handling 4x peak loads
• HR Assistant Platform: Implemented AI-driven HR solutions at Jio Platforms Ltd, streamlining employee support
• Engineered multi-format document extraction pipeline, processing 1200+ files, with OCR & semantic analysis
• Architected Graph RAG pipeline with dynamic agentic chunking, improving the retrieval accuracy by 30%
• Devised modular retrieval pipelines leveraging Qdrant, Graph RAG connectivity & Cross-Encoder refinement
• Leveraged LangChain for multi-turn chats, improving coherence, response clarity, and personalization by 20%

Blockchain-Based Voting Solution