Hi, I'm Ujwal

AIML Engineer

Building from scratch is what I do for fun.

Career Path

A.P Moller Maersk

Sept 2025 - Present

AIML Intern•Bangalore•Expand for Details

•Architected and built an end-to-end distributed entity resolution system combining semantic retrieval, feature engineering, ML-based matching, and cross-encoder reranking, with a fine-tuned cross-encoder trained on augmented data and processed across distributed GPUs in Databricks, improving resolution accuracy from 30% to 75%.
•Implemented Spark ML pipelines to process 350M+ rows, applying feature engineering and hyperparameter tuning to optimize ML-based matching models at scale.
•Leveraged distributed GPUs to efficiently process 100M+ rows using cross-encoder inference for large-scale semantic matching and reranking.
•Optimized Qdrant vector storage to reduce memory requirements by 95% while retaining 97% precision, and built controlled PySpark LLM pipelines for large-scale quality validation and batch inference.
•Built Azure Data Factory pipelines to load 50M+ row data tables from Data Lake into PostgreSQL, implementing indexing strategies and complex data-type handling while reducing data-loading time by 50%.
•Implemented incremental pipelines to ingest 50M+ embeddings into a graph database, enabling hybrid Cypher + vector similarity queries for graph-based retrieval and analysis.
•Developed a graph traversal pipeline using PySpark GraphFrames for efficient connected-component analysis and contributed to an agentic AI system orchestrating database access through tool-based agent workflows.
•Developed geographic address standardization and treatment pipelines across 300K+ addresses to improve consistency and usability of location data.

7thGear AI

May 2025 - Aug 2025

AIML Intern•Bangalore•Expand for Details

•Designed multimodal video processing pipelines combining computer vision, Vision Transformers, NLP, and LLMs to extract and structure business relevant information from multimedia data.
•Created data augmentation pipelines and trained custom CNN models for image classification, improving model accuracy from 60% to 95%.
•Quantized and Fine-tuned the Llama 3.2 3B model using parameter-efficient fine-tuning techniques for domain-specific inference.

LimitlessMinds AI

Jan 2025 - Mar 2025

AI Developer•Delaware•Expand for Details

•Designed data ingestion pipelines to transform unstructured documents into a GraphRAG knowledge base containing structured product information, and developed tools for RAG agents to query the graph and answer customer questions.
•Contributed to a peer-to-peer, session based Voice AI platform using LiveKit and GPT-Realtime, enabling fast, low latency voice AI for answering customer queries.

Research Implementations

LLM Transformer Architecture from Scratch

PyTorch·BPE·RoPE·RMSNorm

Training LLM to Reason with Reinforcement Learning(GRPO)

PyTorch·GRPO·SymPy·SmolLM2-135M

Distilling LLMs: A Reverse-KL Implementation

PyTorch·Reverse KL·Qwen2.5·On-Policy Distillation

Projects

Multi Agent RL Env for Disaster Relief (Meta Hackathon)

OpenEnv·FastAPI·Docker·Python

Agentic AI Self-Healing CI/CD Pipeline

Qwen-2.5-Coder·MCP·GraphDB·GitHub Actions

AdaptIQ — Agentic Employee Onboarding(IISC hack)

Qwen·GPT-OSS·FastAPI·Next.js

Multi Neural Network Mental Health Assessment

Python·TensorFlow·PyTorch·OpenCV

Now

Skills

Languages3 Technologies

ML / Deep Learning10 Technologies

LLM / GenAI17 Technologies

Big Data7 Technologies

Databases4 Technologies

Backend / Systems4 Technologies

Cloud / DevOps7 Technologies

Building

RL Environment to GRPO Fine-Tune LLMBUILDING

Base Inference System for LLM ServingBUILDING

Learning

Reinforcement Learning using openENVLEARNING

GPU Architecture and Data TransferLEARNING

GPU ProfilingLEARNING

Triton KernelsLEARNING

LLM InferenceLEARNING

Activity

GitHub Contributions

ujwal-s-r

2026 Activity

LessMore

Latest Repositories

Loading public repositories...

LinkedIn Feed

Multi-Head Attention GPU Execution

If Multi-head attention splits input across heads for parallel GPU execution... doesn't each head loose data about other heads?

01

Attention Is All You Need From Scratch

As part of me learning to implement DeepLearning from scratch I started with implementing the “Attention Is All You Need” Transformer completely from scratch in PyTorch, raw tensors and matrix multiplication.

02

Fine-Tuning a 7B LLM on GPU

How I (Almost) Choked My GPU While Fine-Tuning a 7B LLM – And How to Avoid It!

03

Forcing CoT vs Teaching Reasoning

Forcing a Model to <think> with (CoT) vs. Teaching a Model to <think> with Reinforcement Learning.

04

Random Weights vs Xavier Weights

Random Weights v/s Xavier Weights. I used to treat weight initialization as a default setting. Random weights, hit run, move on.

05