Nitin Agarwal

Hello All, Here is a talk where I've expressed my perspective on how to place Indian knowledge systems in the modern day context of AI and technological advances.  Some of you may find it interesting -  https://lnkd.in/gbJiA_fw

304

13 Comments

India should build their own LLM. I have done my Masters in AI. My master's thesis focused on generating programming language for modeling like SysML using transformers. I did it on 2 A100 GPUs with 80GB memory. It wasn't as flawless as ChatGPT, but it was showing results. I did it without using the BERT encoder and decoder. There are many open-source Transformer models available which can be fine-tuned. But I created the Embedding Layers, Encoder-Decoder, and the Attention layer from PyTorch from scratch. I ran it for multiple days when I was getting the CUDA out of memory error. I changed the floating point from 32-bit to 16-bit. There were many challenges regarding the data cleaning and gathering. I was reading the DeepSeek paper and found that they are using 8-bit floating point, and they made it work without any impact on accuracy. Using 8-bit floating point is a huge memory saver since you are directly reducing memory consumption by 1/4 (32/4 = 8 bit). They have also ingeniously used Mixture of Experts algo to make LLM more accurate. Then the RLHF magic. More detail regarding the DeepSeek paper I will publish as posts. It's not as hard as people make it out to be. Everyone just assumes that it would be impossible since only companies like Google, OpenAI, and Meta can build it. That's not true. Approach me if anyone wants to invest in this area. Tejeshwi Sharma Aman Goel Aviral Bhatnagar #AI #MachineLearning

339

39 Comments

Fathom-R1-14B, developed by the AI Research team at Fractal, was featured in Mint in a recent roundup of India's open-source AI models - recognized as one of the most downloaded among homegrown releases in May 2025.   Fathom is a powerful 14B parameter reasoning model designed for accuracy, transparency, and adaptability. It can also support a wide range of enterprise-grade applications, while enabling researchers and developers to build responsibly on top of it.   Still early days, but encouraging to see momentum building for open, reasoning-focused models developed in India.   Model: https://lnkd.in/dF6Qxs-d

179

3 Comments

Can you solve this streaming log problem using Python? No pandas. No LLM. Just core Python, smart generators & clean code. Aumkar and I are running a program which is all about implementing solutions focused on #RiskTech. Someone who can think beyond the IDE and can wrangle Python quite well in mission critical production applications is our ideal person to brainstorm with... Think you've got it? DM your solution or drop a GitHub link below You’ve been handed a live production log file that is gigabytes in size — too large to load into memory. Each line in the log file contains a JSON object with a user ID and timestamp of login. Your task: """ Write a Python generator function `get_most_frequent_users` that takes a file path  (of the log file), and yields the top 3 most frequent user IDs **so far**  after reading every 1000 lines. You CANNOT load the entire file into memory. Sample log file lines (each line is a JSON string): {"user_id": "alice", "timestamp": "2025-07-15T10:15:30Z"} {"user_id": "bob", "timestamp": "2025-07-15T10:17:21Z"} Hint: This is a stream-processing problem. """ Expected Output: Every 1000 lines, the generator should yield a tuple like: [("alice", 132), ("bob", 127), ("charlie", 109)]

45

4 Comments

Recently, PW (Physics Wallah) launched Aryabhata 1.0, an AI model built using something called an SLM, which stands for Small Language Model. Unlike big models like GPT-4 or Gemini (called LLMs – Large Language Models) that need a lot of data, power, and memory to run, SLMs are smaller, faster, and easier to use. They are made for specific tasks and give quick answers without using too many resources. This makes them a good fit for education, where speed and accuracy are more important than handling every type of question. While LLMs try to do everything, SLMs are focused — perfect for things like doubt-solving, tutoring, and exam help. Note: A lot of news has surfaced claiming that it is just a wrapper on top of the Qwen model. However,this post is solely focused on how LLM and SLM models differ.

730

8 Comments

📣 On HeatWave, MySQL database and GenAI. Q&A with Nitin Kunal. I talked to Nitin Kunal one of the founding engineers of the #HeatWave project having written its first header file and designed and developed one of its most advanced query optimizers. Currently, He leads the HeatWave server team, responsible for delivering core features of the engine. 👉 Some of the projects they are working on include developing cloud-aware query processing algorithms and applying machine learning to manage databases and optimize data processing. https://lnkd.in/efqsFx-Z #Database #Databases #HeatWave, #MySQL #GenAI #machinelearning #cloud #queryprocessing

60

2 Comments