Day 2

Compete to recreate images through creative prompt iterations. Teams generate visuals using MachineHack AI Playground based on emcee-given prompts. The most accurate and imaginative design wins!

This talk explores the evolution and capabilities of reasoning-focused large language models (LLMs) compared to traditional LLMs trained on next-word prediction. Traditional LLMs excel at in-context learning but struggle with complex reasoning tasks due to their design. New reasoning-optimized models, like Deepseek R1 and OpenAI’s o1, overcome these limitations through techniques like chain-of-thought reasoning and GRPO.

The session highlights practical use cases. Attendees will learn how to adapt prompts and workflows to maximize the potential of reasoning LLMs in coding, planning, medical diagnosis, and beyond.

In this talk, Ayushman will delve into the evolving world of AI agents, focusing on practical techniques and strategies for enhancing their performance, accuracy, and reliability. By examining real-world case studies from diverse sectors—including legal, finance, healthcare, and retail— he will illustrate how Agentic AI systems are reshaping customer experiences and driving measurable business growth. Central to the discussion is a set of targeted approaches and scalable frameworks designed to bolster Large Language Model (LLM)-based agents. These methods promise substantial improvements across various metrics, empowering AI agents to deliver robust, value-driven solutions. The talk will also highlight the ever-growing significance of LLM agents in the technological landscape and their potential to become integral components of future digital experiences. Through these insights, Ayushman aims to provide attendees with actionable strategies and best practices to improve the effectiveness of AI agents in their own organizations, showcasing how the right blend of innovation, customization, and scalability can lead to transformative outcomes.

This research work presents a Generative AI framework for automated digital ad creation, addressing challenges in photorealism and product-background transitions. It proposes a three-step solution tailored to the advertising industry: (a) improved image preprocessing with Canny Edge Detection for precise inpainting, (b) optimized parameter settings for Stable Diffusion ControlNet to generate high-quality images, and (c) a customization module for templates, fonts, and text. The framework streamlines ad creation, enhances visual quality, and reduces costs, offering a practical tool for designers.

This research introduces an agent-driven Text-to-SQL approach using fine-tuned Small Language Models (SLMs) and Retrieval-Augmented Generation (RAG) to simplify enterprise database interactions for non-technical users. It employs specialized agents for tasks like entity recognition, query decomposition, self-correction, and explainable results. Rolled out to 700 users in a pharmaceutical company, the system achieved 87% accuracy, reduced SQL effort by 70%, improved data access by 90%, and cut costs by 50%, enabling scalable, conversational self-service analytics for business users.

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.