Building Scalable AI Systems
from First Principles
About Me
The story behind the code.
I’m a Data Scientist at Fidelity Investments.
I build and deploy production-grade AI systems, with a focus on transformer models, large language models, and inference efficiency. My work sits at the intersection of research and real-world constraints,turning ideas from papers into systems that are reliable, scalable, and cost-aware.
I graduated from IIT Palakkad, where I learned to approach problems from first principles,understanding why something works before deciding how to build it. I stay closely aligned with current developments in AI, regularly experimenting with new architectures, tooling, and workflows.
Alongside building systems, I write extensively about machine learning and create small tools to better visualize and understand how models learn and make decisions.
Professional Experience
Building impactful ML systems at scale, from startup velocity to enterprise reliability.
Fidelity Investments
CurrentData Scientist
Working as a Data Scientist at Fidelity Investments.
Indium (Client: Uber)
Data Scientist
Worked as an external data scientist for Uber, building production-grade AI systems for global earner document processing and onboarding.
- Built a Transformer-based document auto-transcription system for driver and vehicle documents.
- Fine-tuned and deployed LLMs for low-latency, real-time inference in production.
- Automated manual document review workflows, reducing operational effort and cost.
LTIMindtree
Sr. Software Engineer
Worked on applied AI and Generative AI systems for enterprise automation and analytics use cases.
- Built multimodal AI systems to analyze technical videos and generate summaries.
- Developed speech-based natural language interfaces for database querying.
- Implemented RAG-based GenAI chatbots and improved ML model accuracy in production systems.
- Applied transfer learning to achieve high accuracy with limited training data.
Building Systems That Scale
From research implementations to production-grade ML systems serving millions.
Language Identification Model
83.5M-parameter model optimized for source-code classification. Fine-tuned across 25+ programming languages to deliver high-accuracy language detection at scale. Adopted by 350k+ developers worldwide.
Building DETR from Scratch: Object Detection with Transformers
An end-to-end PyTorch implementation of DETR (Detection Transformer), replacing anchor boxes and NMS with learnable object queries and a transformer encoder-decoder. Covers ResNet backbone integration, 2D positional encoding, and Hungarian matching for training supervision.
The Algorithm That Made DETR Possible: Hungarian Loss Explained
A deep dive into the Hungarian algorithm and bipartite matching that powers DETR's set-based object detection. Explains the assignment problem, the matching cost formulation combining classification probability, L1 distance and GIoU, and why auxiliary decoder losses accelerate convergence.
Vision Transformer (ViT) From Scratch
A from-scratch PyTorch implementation of Vision Transformer with configurable hyperparameters and intuitive explanations of self-attention, patch embeddings, and training logic, focused on learning fundamentals and clean engineering. Documented end-to-end on Medium.
Transformer From Scratch
A clean PyTorch implementation of Transformer encoder and decoder from first principles, deeply exploring attention mechanics and causal masking. Guided code and narrative help understand how Transformer blocks actually work.
Distributed Training with PyTorch DDP
A practical deep dive into multi-GPU training using PyTorch DistributedDataParallel (DDP). Explains how gradient synchronization works under the hood, how to scale batch sizes correctly, and how to avoid common pitfalls when moving from single-GPU to distributed training.
Recent Articles
Thoughts on ML systems, engineering practices, and research insights.