I train Deep Neural Networks from the ground up - designing neural network architectures, writing training loops, and building sovereign foundation models from nothing. I don't prefer to fine-tune on top of existing LLMs - I create the recipe from the ground up and train the foundation model with diversified datasets. My work spans AI research - Dense decoders, Mixture of Experts, and Encoder architectures - covering the full lifecycle from pre-training on raw text to post-training with SFT, RLHF, and Chain-of-Thought. I build GenAI products for both research and commercial use. I spend my hours with GPUs, turning compute into intelligence. I write to break down and simplify AI/ML and LLM concepts for the world - making the complex accessible. I also train students and provide technical consultations to industry leaders.
By profession, I am a Senior Technical Account Manager at Amazon - helping top-tier AWS customers architect, operate, and scale in the cloud and AI.
Pet Projects - Research and Use
These are the projects I have built from ground up - each designed to solve a real-world problem. During these projects, I have trained multiple foundation models spanning Dense decoders, Mixture of Experts, and Encoder architectures.
-
Compliance Classifier Router (134M Encoder)
A BERT-based encoder model that classifies incoming queries by complexity and PII sensitivity, routing them to the appropriate LLM tier. 99.2% accuracy, 7ms latency, deployed on ECS with full CloudFormation IaC. Enforces data residency for Indian financial services.
-
Indian Multi-Language Model (3B)
Pre-trained on 6 Indian languages - Hindi, Bengali, Tamil, Telugu, Marathi, Kannada. Uses a 64K multilingual tokenizer trained across all languages. Designed for a country where most AI tools only work in English.
-
Sanskrit Translation Model (3B-7B)
Bilingual English-Sanskrit model trained on Vedic texts, epics, and modern Sanskrit. Handles translation, grammar correction, and transliteration. Preserving an ancient language through modern architecture.
-
Drone Programming - Instruction Set Generator (<1B)
A decoder model that converts English prompts into executable drone flight code. Non-programmers describe a mission and the model generates the SDK commands. Includes a safety validator that checks altitude limits and whitelisted functions before execution.
Featured Writing
-
What Does 70B Parameters Actually Mean?
Breaking down what 70 billion parameters actually represent, where the number comes from, and when bigger genuinely matters versus when it's expensive overkill.
-
Can AI Develop Intelligence On Its Own?
Exploring whether a pre-trained model left alone can develop intelligence, what AGI actually requires, and why the real race is between humans, not machines.
Research
In 2026, I am researching on - Deep Neural Networks and hidden layers, classifier-gated routing for cost optimization and performance, sovereign language models for Indian languages (Sanskrit, Hindi, Tamil, Bengali, Kannada), purpose-built small and tiny models for efficiency, data residency and compliance by design, and the intersection of compliance and AI architecture.
Videos
Papers & Publications
Academic
I hold a BSc in Physics (Honours) with Mathemetics, graduated with with first class grade - building the mathematical foundation that now drives my neural network research. I earned a Bachelor of Engineering degree in Computer Science with 8.4 CGPA in 2011, giving me the systems thinking to architect and train models at scale. I am a proud alumnus of the Indian Institute of Management (IIM) Calcutta - one of Asia's premier business schools. Currently pursuing a Master of Arts in Philosophy, exploring the deeper questions of intelligence, cognition, and what it means for machines to "learn."
Work Experience
