AI Meetups by Deep Tech Stars

In this hands-on session, you’ll learn about Sarvam’s latest models, Sarvam‑105B, Saaras V3 (STT), and Bulbul V3 (TTS). You’ll see how these models handle multilingual reasoning, noisy voice inputs in 22 Indian languages, and expressive text‑to‑speech outputs. You will also get introduced to BharatML Stack, the open source production-grade ML infrastructure platform from Meesho.

Bring your laptop to build and experiment live. Whether you’re working on customer support, voice systems, workflow automation, or India‑first products, you’ll walk away with concrete patterns for integrating these stacks into real‑world applications.

AGENDA
11:30 AM – Welcome note
11:45 AM – Walk-through of BharatML Stack
1:00 PM – Lunch and networking
2:00 PM – Walk-through of Sarvam‑105B – Sarvam’s Sovereign LLM
2:15 PM – Walk-through of Saaras V3 - Sarvam’s Speech to Text SOTA model
2:30 PM – Walk-through of Bulbul V3 – Turning your agent’s responses into natural voice
2:45 PM – Get hands-on with Sarvam models and best practices
4:00 PM – Conclusion

This hands-on session dives deep into NVIDIA’s newly released Nemotron 3 Super, a powerful 120B open hybrid Mamba-Transformer Mixture-of-Experts (MoE) model built for complex reasoning, long-context analysis, and autonomous problem-solving. You will explore key architectural innovations like Latent MoE and Multi-Token Prediction, and learn how to build, optimize, and deploy real-world AI systems using NVIDIA’s open weights, datasets, and deployment recipes.

WHO SHOULD ATTEND
AI/ML Engineers & Researchers: Professionals building advanced LLMs and specialized reasoning models.
Generative AI Developers: Developers creating complex, multi-agent AI applications (such as software development or cybersecurity triaging agents).
Data Scientists: Practitioners interested in state-of-the-art training techniques like Native NVFP4 pretraining and multi-environment reinforcement learning (RL).
DevOps & MLOps Engineers: Professionals looking to deploy high-throughput, low-latency AI models using vLLM, SGLang, or NVIDIA TensorRT-LLM.

AGENDA
10:00 AM – 10:30 AM: NVIDIA Nemotron strategy – open datasets, weights, frameworks, and recipes
10:30 AM – 11:00 AM: Deep dive into Nemotron 3 Super architecture design
11:00 AM – 12:15 PM: Efficient serving and deployment using vLLM, SGLang, and TensorRT-LLM Cookbooks
12:15 AM – 1:30 PM: Building agentic workflows using Nemotron models

KEY LEARNINGS
By the end of the workshop, attendees will be able to:

• Understand Advanced Architectures: Grasp the mechanics behind the Hybrid Mamba-Transformer backbone, Latent MoE (which calls 4x as many experts for the same compute cost), and Multi-Token Prediction (MTP) for built-in speculative decoding.
• Navigate the Training Pipeline: Understand how the model achieves stability and accuracy through Native NVFP4 pretraining and trajectory-based reinforcement learning across diverse environments (using NeMo Gym and NeMo RL).
• Implement the "Super + Nano" Pattern: Learn how to architect multi-agent workflows that smartly route simple tasks to Nemotron 3 Nano and complex planning/reasoning tasks to Nemotron 3 Super.
• Deploy and Fine-Tune: Utilize NVIDIA’s open resources and deployment cookbooks (vLLM, SGLang, TensorRT-LLM) to customize and run the model on your own infrastructure.

This hands-on session dives deep into NVIDIA’s newly released Nemotron 3 Super, a powerful 120B open hybrid Mamba-Transformer Mixture-of-Experts (MoE) model built for complex reasoning, long-context analysis, and autonomous problem-solving. You will explore key architectural innovations like Latent MoE and Multi-Token Prediction, and learn how to build, optimize, and deploy real-world AI systems using NVIDIA’s open weights, datasets, and deployment recipes.

WHO SHOULD ATTEND
AI/ML Engineers & Researchers: Professionals building advanced LLMs and specialized reasoning models.
Generative AI Developers: Developers creating complex, multi-agent AI applications (such as software development or cybersecurity triaging agents).
Data Scientists: Practitioners interested in state-of-the-art training techniques like Native NVFP4 pretraining and multi-environment reinforcement learning (RL).
DevOps & MLOps Engineers: Professionals looking to deploy high-throughput, low-latency AI models using vLLM, SGLang, or NVIDIA TensorRT-LLM.

AGENDA
10:00 AM – 10:30 AM: NVIDIA Nemotron strategy – open datasets, weights, frameworks, and recipes
10:30 AM – 11:00 AM: Deep dive into Nemotron 3 Super architecture design
11:00 AM – 12:15 PM: Efficient serving and deployment using vLLM, SGLang, and TensorRT-LLM Cookbooks
12:15 AM – 1:30 PM: Building agentic workflows using Nemotron models
1:30 PM - 2:15 PM: Lunch
2:15 PM - 2:30 PM: Talk by Partner
2:30 PM - 2:45 PM: Paper/project presentation 1
2:30 PM - 2:45 PM: Paper/project presentation 2
2:45 PM - 3:30 PM: Talk by Partner
3:30 PM: Networking & Wrap-Up

KEY LEARNINGS
By the end of the workshop, attendees will be able to:

• Understand Advanced Architectures: Grasp the mechanics behind the Hybrid Mamba-Transformer backbone, Latent MoE (which calls 4x as many experts for the same compute cost), and Multi-Token Prediction (MTP) for built-in speculative decoding.
• Navigate the Training Pipeline: Understand how the model achieves stability and accuracy through Native NVFP4 pretraining and trajectory-based reinforcement learning across diverse environments (using NeMo Gym and NeMo RL).
• Implement the "Super + Nano" Pattern: Learn how to architect multi-agent workflows that smartly route simple tasks to Nemotron 3 Nano and complex planning/reasoning tasks to Nemotron 3 Super.
• Deploy and Fine-Tune: Utilize NVIDIA’s open resources and deployment cookbooks (vLLM, SGLang, TensorRT-LLM) to customize and run the model on your own infrastructure.