AIM, in partnership with NVIDIA, successfully conducted the fourth edition of DevPalooza, the prestigious developer meetup series organised under the AI Forum Community at NVIDIA’s Bengaluru Office last week.
With around 50+ developers on board, this edition of DevPalooza, themed ‘Building the Future: Hands-On with Generative AI,’ highlighted the need to understand techniques and build real-world applications of generative AI.
Designed for AI professionals and practitioners working with AI workloads on GPU computing, DevPalooza 4.0 offered an immersive experience with practical sessions and collaborative discussions.
Attendees participated in hands-on workshops with industry experts like Usha Rengaraju, AI consultant, Corporate trainer & Influencer; Suvojit Hore, Senior AI Engineer, EXL; and Shreyans Dhankhar, Senior Solutions Architect, NVIDIA.
Kavita Aroor, Head Developer and Startup Marketing for South Asia at NVIDIA, also spoke at the meetup, highlighting the NVIDIA developer ecosystem and its range of tools for building generative AI applications.
NVIDIA is all set to host its AI Summit 2024, from October 23 to 25 at the Jio World Convention Centre in Mumbai, India. The summit will feature three days of presentations, hands-on workshops, and networking opportunities, bringing together industry experts to explore the latest advancements in AI.
Register now with this referral code NV0ZXHTAE to save 25% off your pass and not miss out on Jensen Huang’s Keynote.
Key Takeaways from DevPalooza 4.0
- Efficient AI Application Development: The hands-on demo of NVIDIA’s NIM with LangChain showcased how entire applications can be built with just 15-20 lines of code, emphasising the efficiency of these tools in developing generative AI applications.
- Addressing Hallucination in LLMs with RAG: Usha introduced RAG and GraphRAG as solutions to hallucinations in LLMs, highlighting how domain-specific knowledge stores improve response accuracy, while also explaining how to build the whole pipeline.
- AI Twins for Multiple Use Cases: Suvojit explained how AI twins can replicate speech patterns and behaviour, offering practical uses in customer service, healthcare, and even mental health, backed by a live demo.
- Building an AI Twin: Suvojit detailed the architecture for creating AI twins, including using NVIDIA Parakeet for speech-to-text and HiFiGAN for text-to-speech conversion, enabling real-time, human-like AI interactions.
- Cost Benefits of On-Premise LLM Deployment: Shreyans compared the costs of on-premise LLM deployment with third-party services like OpenAI, demonstrating that NVIDIA’s solutions offer significant savings and better control over latency and scalability.
- NVIDIA’s Open-Source Ecosystem: Shreyans highlighted NVIDIA’s range of open-source SDKs, such as MONAI and NeMo, which are free for developers to explore and integrate into their projects, catering to healthcare, generative AI, and also building a multimodal pipeline.
Building GenAI Applications Using NVIDIA’s NIM
DevPalooza 4.0 began with Usha’s session on ‘Building GenAI applications using NVIDIA’s NIM’, which became one of the highlights of the meetup because of her insights and the questions developers asked. Providing a brief introduction to the concepts of RAG and GraphRAG, and focusing on the practical applications of generative AI, Usha emphasised the challenges of hallucination in LLMs. She explained that RAG addresses this issue by allowing users to create domain-specific knowledge stores.
“RAG was one of the techniques proposed to solve the problem of hallucinations,” she noted, illustrating how the approach enhances the accuracy of AI responses.
The hands-on demonstration showcased the integration of NVIDIA’s NIM (NVIDIA Inference Microservices) with LangChain, a high-level orchestration framework. “We are going to build an entire application with 15-20 lines of code,” Usha said, underscoring the efficiency of these tools.
A significant part of the session was dedicated to differentiating between RAG and GraphRAG. Usha stated, “GraphRAG will be able to capture the nuances of the relationships between different entities,” making it especially beneficial for complex, interconnected data environments.
Throughout the session, Usha encouraged audience interaction and provided insights into using these technologies for various applications, including chatbots and question-answering systems.
Building an AI Twin
In his session, Suvojit from EXL humorously introduced a recorded message from Elon Musk, explaining that an AI twin mimics a person’s speech patterns, tone, and behaviour, effectively acting as a digital replica of that individual. He noted that AI twins can serve several practical applications, such as customer support agents or even customer personas used to train agents.
Suvojit then offered a detailed explanation on building such AI twins, guiding attendees through a hands-on demo and emphasising the key components required for its development.
He outlined the architecture required for building an AI twin, utilising tools such as NVIDIA’s Parakeet for speech-to-text conversion and LLMs like Mistral, powered by the NIM API. He described how audio from a user is transcribed, processed by an LLM, and then converted back into audio using HiFiGAN for text-to-speech, which is subsequently lip-synced to a video via the WAV2LIFT model.
The use cases he highlighted ranged from automated handling of less critical meetings via an AI twin and customer service automation, to healthcare assistants for routine medical consultations. AI twins could even play a role in mental health coaching and education by providing interactive, personalised learning or therapy sessions.
To help attendees visualise the process, Suvojit presented an end-to-end demo, showing how pre-recorded video footage can be synced with audio generated from LLM responses, creating a realistic interaction with the AI twin.
From LLM Inference to Multimodal RAG: A Hands-On Journey from Theory to Application
The last session with Shreyans was about the whole pipeline of building from inference to multimodal RAG. His session kicked off by explaining that NVIDIA offers a range of tools beyond GPUs, such as CPUs, DPUs, and a host of software solutions that accelerate deep learning tasks. “We developed CUDA, and many of you working in deep learning may have used it. All of this comes from NVIDIA,” he said.
Shreyans also highlighted NVIDIA’s AI platforms, including NVIDIA AI and Omniverse, which cater to AI activities and digital twin development, respectively. He further spoke about the company’s open-source offerings, showcasing various SDKs like MONAI for healthcare and NeMo for generative AI. “99% of our SDKs are free and open source,” he noted, inviting attendees to explore these tools on GitHub.
One of the key takeaways from the session was the detailed comparison of deploying LLMs on-premise versus using third-party services like OpenAI. “When you compare costs, NVIDIA’s on-prem deployment can offer significant savings. For instance, we calculated that per million tokens, OpenAI charges nearly four times more than what we offer with a single GPU setup,” Shreyans said.
He also demonstrated how NVIDIA’s solutions allow for greater control over latency and scalability, making them more feasible for enterprises looking to optimise costs.
During the hands-on section, Shreyans walked the participants through various benchmarks and deployment strategies. He explained developers can easily benchmark and adjust deployments according to their service level agreements, showcasing the flexibility of NVIDIA’s inference and multimodal capabilities.
Wrapping Up
Overall, DevPalooza 4.0 was a blast for developers and provided useful insights and guidelines on building AI applications using NVIDIA tools.
Stay tuned for the next edition of DevPalooza! Join our AI Forum Community to stay updated.
The post NVIDIA and AIM Power Up Developers with GenAI Masterclass at DevPalooza 4.0 appeared first on AIM.
