Jitesh Kumar Dewangan

Agentic AI ChatBot: Developed an internal validation dashboard chatbot leveraging Agentic AI and MCP server capabilities, powered by LangChain, a GPT-4o model, and RAG (Retrieval Augmented Generation). This solution significantly increased user productivity, reduced manual effort, and provided real-time critical condition updates and notifications to platform owners. Intel Defect System: Leading the development of Machine Learning model for a similar defect recommendation system, significantly enhancing identification and resolution efficiency. Use Case Prediction System: Developing and deploying an ML-driven use case prediction system to proactively identify potential defect scenarios and improve product reliability.

Large-Scale Speech-to-Text Systems: Designed and developed robust ASR systems and language models powering GalaxyAI features like Live Call Translate, Interpreter, and Call Assist, deployed into production with high reliability across millions of devices. Hybrid Transformer-Based ASR Architecture: Led the implementation of a Transformer-based cascaded architecture supporting both streaming and non-streaming ASR modes. Achieved 2% improvement in recognition accuracy and reduced first-pass latency through architectural and efficiency-focused enhancements. Scalable Data & Training Pipelines: Built and productionized scalable ETL and ML pipelines, enabling efficient data ingestion, transformation, and training workflows. Enabled consistent delivery of both real-time and offline model artifacts. Multilingual Model Adaptation with LoRA: Applied Low-Rank Adaptation (LoRA) to fine-tune multilingual ASR models, achieving 10% relative accuracy improvement across diverse linguistic domains while minimizing training overhead. Model Efficiency via Knowledge Distillation: Optimized large ASR models into lightweight on-device variants using Knowledge Distillation, accelerating training convergence and enabling real-time inference with minimal degradation in accuracy.

Warehouse Automation: Developed sophisticated optimization algorithms using Erlang to enhance storage management and throughput efficiency for GreyOrange's global warehouse automation solutions, serving clients worldwide. Robot Operation Flow: Enhanced put fulfillment operations by designing and implementing advanced heuristic-based algorithms to optimize rack selection and storage utilization, significantly improving overall warehouse throughput and operational efficiency.

Name Entity Correction: Designed and implemented an on-device NEC algorithm to accurately correct app and contact names, enhancing end-to-end ASR performance and achieving a substantial 7-8% improvement in accuracy. Device Validation Tool: Created a comprehensive validation tool to meticulously assess system partition memory, app versions, and layout configurations, streamlining processes and reducing manual effort by an impactful 23%.