Alex Mashrabov

In 5 days, we’re going to showcase how we're solving one of voice AI's biggest challenges - bridging the gap between what customers say and how they say it. Gnani.ai teams are going to be attending NVIDIA's GTC Conference in San Jose, California on March 17 to exhibit our speech-to-speech AI models.  As one of the first few companies building these models globally, our partnership with NVIDIA to use their innovative platform to solve critical voice AI challenges is crucial and groundbreaking.  Why we’ve been invited to speak: The most challenging piece about voice AI agents is that when you start using them, you realize the software pipeline only looks into the text behind the conversations. As we were implementing voice AI agents for multiple customers, we realized we were missing essential elements: the pitch, the tone, and the way customers react to certain words the bot was saying. So, in addition to solving the classic LLM problems of latency and hallucinations, which continue to be big challenges, we also needed to make the voice AI agents’ conversations more authentic. To do this, we decided to build speech-to-speech AI models for real-time performance, from voice recognition to response generation. Our software stack makes customer interactions fast, natural, and context-aware. This idea resonated with NVIDIA. That’s why we’re building these models on broader NVIDIA architecture. What we’ll be presenting:    At GTC, we’ll be sharing the specific innovations we've invented and what this really means for not just voice AI agents, but the voice AI industry as a whole. We’ll also reveal how our solution: - Is trained on over 14 million hours of multilingual conversational data - Has out-of-the-box support for 14 languages - Is equipped with accurate, context-rich responses that adapt to real-time interactions - Transforms raw data into strategic customer intelligence - Provides insights that elevate enterprise decision-making   - Incorporates industry-leading software like NVIDIA TensorRT-LLM, NVIDIA Triton and NVIDIA Riva The key features we're presenting are about how we’re able to merge multiple elements in the pipeline of a voice AI agent so that we can reduce the latency to 250ms--making interactions human-like. We’ll share the algorithms we used and the pipeline we used to make this possible, as well as the broader NVIDIA architecture that made it possible to design these voice agent flows for customers. It’s going to be great. See you there! Ganesh Gopalan Bharath Shankar Thoshith S Avinash Benki John Philip NVIDIA NVIDIA GTC

129

10 Comments

As Jensen mentioned with Brad Gerstner & Bill Gurley, something that few people know is that NVIDIA is becoming the American open-source leader in AI, with over 400 contributions of models, datasets and apps on Hugging Face in the 12 months. And I have a feeling they're just getting started!

764

29 Comments

🚀 Big news for AI agent builders: Agent Reinforcement Trainer (ART) now integrates seamlessly with 🦜🔗 LangGraph! If you’ve ever built multi-step, ReAct-style agents, you know the pain: endless prompt tweaking, fragile reasoning, and lots of manual trial-and-error. With this new integration, you can now train your LangGraph agents with reinforcement learning — making them smarter, more adaptive, and more efficient over time. ✨ What this means in practice: * Agents learn to reason through multi-step problems without hand-holding * They improve tool usage automatically (no more prompt spaghetti) * They make smarter decisions in complex workflows * Training is scalable and robust, powered by ART’s reinforcement engine And the best part? It’s a drop-in upgrade: * Swap in ART’s init_chat_model and wrap_rollout * All interactions, tool calls, and reasoning chains get logged automatically * Plug straight into ART’s RULER reward function for training — no custom rewards needed 💡 Imagine your LangGraph prototypes evolving into production-ready agents that actually get better with use. That’s the power of combining LangGraph’s agent framework with ART’s reinforcement learning. 👉 Full docs & examples here: https://lnkd.in/gwWj-xxx Can’t wait to see what the community builds with this. If you’re experimenting with multi-step AI agents, give your agents the ability to learn from experience!

1,889

80 Comments

My podcast episode with Art Dicker on The SoCal Edge is now live. 🎙️ We dove deep into my journey from neuroscience and brain-machine interfaces to startups and co-founding Log10, and how we're tackling one of AI's biggest challenges: making it trustworthy enough for high-stakes domains like healthcare and law. Some key points we explored: 🎯 Why accuracy isn't just important—it's everything when AI touches critical workflows ⚡ How Everest, our flagship product, is revolutionizing medical documentation with its focus on accuracy 🚀 The pivotal moments in AI history I've witnessed, from the AlexNet+GPU moment of 2012 to today's data-rich landscape 🛤️ My path from Qualcomm to co-founding Nervana Systems (acquired by Intel) to now building Log10 The future of AI isn't just about more powerful models—it's about AI systems that understand causality and can be trusted in enterprise applications where there's no room for error. For aspiring entrepreneurs: Whether you're pursuing deep research or identifying unmet needs in everyday workflows, the opportunities to build impactful AI solutions have never been greater. Thanks Art Dicker for the engaging conversation! Check out the full episode to hear more about what it takes to make AI truly enterprise-ready. Spotify: https://lnkd.in/dWHE-67f Apple podcasts: https://lnkd.in/dMDPe8_p #AI #MachineLearning #Healthcare #Innovation #Entrepreneurship #Log10

48

1 Comment

Every breakthrough comes with hidden limitations. Right now, embeddings—the foundation of modern RAG systems—stands close to a ceiling. While embeddings are excellent at representing data in vectors they fall short in capturing rich, interconnected relationships—the kind of nuanced, human-like understanding we need for next-gen AI. Knowledge graphs solve this. They preserve information the way humans process it—as interconnected concepts rather than isolated data points. They’re perfect for tackling what embeddings miss: multi-hop reasoning, hierarchical structures and temporal logic. 2025 could be the year knowledge graphs go mainstream, reshaping how AI systems learn and reason. If you’re in AI or data, this is your signal: start exploring knowledge graphs now. What are you betting on in 2025?

29

10 Comments

SiMa.ai’s $85M Raise Marks a New Chapter for Physical AI SIMA AI, a startup focused on edge inference, just secured an $85 million Series C led by Maverick Ventures, with participation from Lockheed Martin and Amplify Partners. But the number isn’t what matters. The thesis is. SIMA is building a fully integrated chip + software stack that enables real-time AI processing on devices that don’t have the luxury of hyperscaler latency budgets. In other words, the future of AI isn’t just chatbots and copilots. It’s the physical world, where inference must be deployed without cloud contracts, without power-hungry GPUs, and without sacrificing latency. CEO Krishna Rangasayee put it bluntly: “Inference shouldn’t need 500 watts and a cloud contract.” As AI moves beyond screens and into machines, companies like SIMA are not just enabling performance—they’re defining the terms of where and how intelligence lives. The future may be foundation models. But increasingly, it also looks like edge-native, domain-specific, silicon-optimized intelligence. And that, perhaps, is the quieter revolution unfolding. #ai #inference #cloud #physicalAI

74

2 Comments