HALL 3 - Tech Talks / Paper Presentations

In this session, I’ll dive deep into the technical challenges and solutions in building enterprise-grade speech AI systems that power content globalization. We’ll explore how Dubverse has developed and scaled speech foundation models, including text-to-speech, speech-to-speech and speech-to-text technologies. I’ll share insights on optimizing model throughput, handling low-resource languages, and building robust APIs. The talk will cover practical approaches to solving critical problems in voice technology, making it relevant for developers working on speech, audio, or multilingual applications.

This research work addresses the critical challenge of on-shelf availability in retail, caused by delays, misplaced inventory, and phantom stock issues. It introduces a scalable AI-powered solution deployed across North American retailers, leveraging advanced ML models (e.g., ARIMA, XGBoost, neural networks) and distributed cloud processing. The system detects disruptions through an adaptive alert framework, enabling targeted resolutions and continuous improvement. Insights into inventory metrics support strategic interventions and mitigate upstream supply chain issues. This approach transforms inventory management, enhancing revenue and operational performance.

This research work presents an advanced solution to streamline the tendering process for engineers by automating Single Line Diagram (SLD) and technical specification processing. Traditional manual methods are inefficient and error-prone. The proposed system integrates OCR with a rule-based engine to extract critical SLD information, including component types and connections. Additionally, an LLM-based approach processes technical specifications, extracting data like requirements and compliance standards. This automation reduces manual effort, enhances accuracy, and accelerates tender responses, enabling engineers to prioritize strategic tasks.

This research work explores transitioning from third-party to in-house open-source OCR solutions to cut costs, improve data security, and enhance flexibility. The solution integrates barcode/QR code detection and supports 1D and 2D formats. It powers downstream applications like Retrieval-Augmented Generation (RAG) and text summarization, ensuring efficient processing. Standardized APIs reduce development time and integration complexity. Key strategies include framework selection, image optimization, and multilingual support. The approach improves cost-efficiency, privacy, and scalability while enabling advanced AI applications.

This research explores AI Twins, lifelike digital avatars enabling real-time interaction through advanced AI-driven speech synthesis, expressions, and contextual understanding. It compares two implementations: a closed-source system utilizing tools like Synthesia and Eleven Labs for high-quality video and audio, and an open-source alternative with Nvidia Riva, Nemotron LLM, Wav2Lip, and Sadtalker for speech and lip-sync. Both use Retrieval-Augmented Generation (RAG) and AI Agents for enhanced domain intelligence, offering applications in personalized education, training, customer support, and virtual companionship.

This research introduces a unified platform for fine-tuning multiple open-source AI models on small datasets, optimizing resource usage and ensuring data privacy. The approach employs Parameter-Efficient Fine-Tuning (PEFT) with QLoRA, dynamic hyperparameter tuning, and stratified random sampling for efficient performance. It automates data preprocessing, prompt adjustment, and compliance with responsible AI practices. By evaluating models on metrics like accuracy, F1 score, and perplexity, the platform provides actionable insights, reducing costs and time while enabling businesses to select the best model for their needs.

World has moved from drag and drop timeline based editors. Text to video is a paradigm which makes video creation and distribution a verbal skill. This takes this confined ability to the hands of every person who can think creatively and has a story to tell. But chatting with a video creation & editing agent has its own set of challenges, especially for the user asking for high content control as well as ease of creation. Existence of 1000+ languages poses another dimension in these challenges. BingeClip AI is configuring a set of technologies which include Agentic RAGs, Automated document workflows, Proprietary video search models, autonomous agents for editing operations and model compression. The combination creates a seamless application layer which enables user interaction, hence video creation on the move, combining authenticity of UGC with creativity & Automation of AI.

This research introduces an innovative taxpayer risk prediction approach leveraging Large Language Models (LLMs). Taxpayer data is transformed into detailed natural-language profiles, enabling fine-tuning of pre-trained LLMs for precise and nuanced risk assessment. Unlike traditional methods, this approach identifies hidden patterns, offering deeper insights into taxpayer behavior. Preliminary results show improved accuracy and decision-making capabilities for financial departments. The research emphasizes the transformative potential of LLMs in modernizing fiscal oversight and ensuring effective governance amid evolving financial landscapes and taxpayer risks.

This research details the creation of an innovative biomedical chatbot aimed at improving patient question-answering. It employs a mixture of expert models—fine-tuning one LLaMA3 on patient-doctor interactions and another on general biomedical datasets—and combines them via a router. A vector database supports Retrieval-Augmented Generation (RAG) for accurate, context-aware responses. Rigorous evaluation across biomedical datasets highlights its ability to enhance patient engagement and healthcare delivery, showcasing the transformative potential of generative AI in medicine.

This research presents a transformer-based approach for Next Basket Recommendation (NBR) in grocery retail, focusing on predicting new, previously unpurchased products. Using a modified T5 architecture, the model captures item-to-item correlations across sequential shopping baskets, framing NBR as a sequence-to-sequence task. Extensive evaluations on the Dunnhumby dataset show that the model outperforms traditional collaborative filtering methods, achieving high precision and ranking effectiveness. This approach demonstrates the potential of transformers for personalized and novel product recommendations.