Swapan Rajdev

I've seen this pattern in every AI app I've worked on. Demos work great, but then real users expose the cracks. Change one prompt and something else breaks. Users and clients start losing faith in AI. The answer to most of these problems? - Strong, clear evaluation metrics that you track regularly. - It gives you the confidence and peace of mind of knowing exactly how your app is performing in production. I wrote a guide on the scorecard I use for all the apps I ship and how to implement it. Link to the full implementation guide in the comments.

Simple stuff like this makes me productive and a bit excited because I feel like I just found a smarter way to work :) Told ChatGPT to go through all the references in a research paper I am currently studying and turn them into a table with links. No more googling each one manually. Note: had to tell it to scrape 10-20 at a time, then ask it to merge them all.

Pretraining gives language models their knowledge, and post-training gives them purpose! But with so many techniques—fine-tuning, instruction tuning, RLHF, test-time reasoning—it’s easy to get lost in the maze. I recently came across a fantastic paper—LLM Post-Training: A Deep Dive into Reasoning LLMs—that mapped out this space beautifully. So I decided to break it all down into something easier to digest. 📘 In this post, I’ve: - Categorized key post-training methods (Fine-Tuning, RL, Test-Time Scaling) - Summarized strengths, challenges, and real-world use cases - Shared clean, visual tables to help you pick the right technique for your application This reference might save you a lot of time if you’re building AI products or working with LLMs in production.

For 30 years, I’ve witnessed the power of a great special educator—because my mother is one. The right guidance doesn’t just help children learn; it transforms entire families, especially for differently abled children. But in India, the reality is stark: 📌 30 million children are estimated to have Specific Learning Disabilities (SLD). 📌 Yet, only 8 million have been officially identified—leaving 22 million children and families unaware. 📌 Even for the identified cases, we have just 1.6 lakh special educators—meeting only 25-35% of the actual need. This massive gap in special education is a challenge—but also an opportunity. Given where AI stands today, I truly believe we can bridge this gap and create a lasting impact. That’s why I started building an AI-powered special educator—to bring personalized, accessible, and scalable support to children, parents, and educators. 🚀 Super excited to share that on the 1-year anniversary of IndiaAI Mission, we were selected as one of 30 companies (from 900+ applicants) to receive a grant from the IndiaAI Application Innovation Challenge. This recognition strengthens our mission to revolutionize special education in India. If this vision resonates with you, I’d love to connect. Whether you're an educator, technologist, policymaker, or just someone who cares deeply about making a difference—we need you. Let’s build the future of special education together. #AIForGood #InclusiveEducation #SpecialEducation #IndiaAI #Accessibility

The AI Stack: A Framework for Every Builder AI is evolving at lightning speed, but you might be playing the wrong game if you don’t understand the AI stack. After years of working in AI, one thing is clear: many companies don’t truly know where they compete, what they own, or how they differentiate. Some think they’re building defensible AI businesses—but in reality, they’re just integrating someone else’s technology. That’s why I wrote this breakdown. It’s not just a framework; it’s a practical lens to help founders, builders, and investors pinpoint where real value is created in AI. If you're building in AI, this will help you: ✅ Understand the five key layers of the AI stack ✅ Move faster by knowing where to focus and where to partner ✅ Identify whether you're creating value or relying on others This might seem obvious, but getting this right is the difference between leading and getting left behind.

With NVIDIA’s exciting GPU launches yesterday, the spotlight on AI hardware has never been brighter. But amidst the buzz, it’s easy to get lost in terms like cores, FLOPS, memory, and bandwidth. What do they really mean for AI, especially for training and deploying Large Language Models (LLMs)? 💡 Inspired by the latest developments, I wrote a blog to break down GPU components and their critical role in the performance of LLMs. From understanding floating-point precision to decoding memory bandwidth, I’ve dived into what makes GPUs the powerhouse behind modern AI. If you’re curious about how GPUs fuel the future of AI—or want to make sense of yesterday’s announcements—this blog is for you!

I’m excited to share my latest blog exploring DeepSeek-V3, a groundbreaking open-source AI model that is setting new standards in performance, efficiency, and accessibility. DeepSeek-V3 rivals industry titans like GPT-4o and Claude 3.5 Sonnet while being trained at a record-low cost of $5.576 million—a feat we delve into, uncovering the innovations that made it possible. Highlights from the blog: ✅ Architectural Innovations: Cutting-edge Mixture-of-Experts (MoE) framework, Multi-head Latent Attention (MLA), and auxiliary-loss-free load balancing for optimized performance. ✅ Training Methodology: Learn how the use of FP8 mixed precision and the custom HAI-LLM framework enabled cost-efficient training on a massive 14.8 trillion-token dataset. ✅ Benchmark Performance: Outperforms Llama 3.1 405B and Qwen while matching GPT-4o and Claude 3.5 Sonnet in reasoning, mathematics, and real-world use cases. ✅ Unparalleled Efficiency: Offers unmatched inference speed and is cost-effective to host, making it ideal for developers and enterprises alike. ✅ Use Cases: From advanced reasoning and multi-modal capabilities to client-facing applications, DeepSeek-V3 is a game-changer for AI-driven solutions. 🌟 Whether you’re an AI enthusiast, a developer, or a leader exploring AI for your organization, this blog unpacks how DeepSeek-V3 is reshaping the landscape of open-source AI.

BFCL benchmarks on generic use cases. If we consider enterprise use cases, the performance can be even worse. - Basic parameter identification failures - Rule adherence breakdowns - Consistent predictability issues To make function calling work in production, we've had to implement lots of hacks and manual interventions. It often feels like patching something that should be core functionality.

Function calling is the backbone of intelligent agent architectures. However, the current capabilities of LLM are nowhere close to where we need them to be. Looking at the Berkeley Function Calling Leaderboard, the landscape is sobering. Top models are achieving only 68.5% accuracy in generic use cases. That's not enterprise-grade reliability. 📊

The "Pricing Assistant GPT" button on @videodb_io's pricing page is such a brilliant idea. I was super confused by the pricing but was able to make sense of it all after chatting with their GPT Great use of @OpenAI's GPTs