Shrirang Mahajan

Full-Stack AI Developer & Machine Learning Engineer — "In search of Global Maximum..."

I am a Computer Engineer from Pune, India. I am passionate about building AI products and solving real-world problems. I have experience building scalable and robust software systems, and I am always eager to learn new technologies and work on challenging projects.

Education

Bachelor's Degree in Computer Engineering (Minor: Data Science) from Savitribai Phule Pune University, CGPA 9.14 / 10.

Skills

Python, C++, CUDA, JavaScript, SQL, PyTorch, Triton, TensorFlow, LangChain, Scikit-learn, OpenCV, FastAPI, Streamlit, MLFlow, PySpark, Git, Azure Databricks, Docker, Hugging Face, AWS, PostgreSQL.

Experience

Machine Learning Engineer II — Skylark Labs

November 2025 — Present

Developed a production-grade visual tracking system featuring a 7D Kalman filter for multi-object tracking and predictive PiP stabilization. Built an offline standalone tracker with Camera Motion Compensation using sparse optical flow and affine estimation, handling complex re-identification and dynamic state switches.
Optimized inference by deploying a TensorRT-optimized model on NVIDIA Triton Inference Server, doubling throughput from 3 FPS to 6–7 FPS via profiling and optimized post-processing logic over gRPC.
Architected and developed Drishti, a distributed annotation system processing video datasets with GPU-accelerated YOLO inference on Celery workers, automated quality control, smart work distribution, and real-time batch processing, deployed as a full-stack solution (Django + React + PostgreSQL + Docker) with Cloudflare R2 integration, improving annotation throughput by 3x.
Worked on continuous training and validation of a YOLO-based object detection model, including hyperparameter tuning and false-positive analysis.

Data Scientist — Emergys (Ellicium)

July 2024 — October 2025

Designed and implemented, as an individual contributor, a comprehensive dashboard for detecting price outliers in product flavor data, incorporating clustering analysis, supplier concentration metrics, and portfolio optimization techniques.
Built a promo-sales forecasting engine using Gradient Boosting (R² 0.93) on Azure Databricks; automated optimization using a TPE-based recommender to suggest an optimal promotional strategy.
Led the development of an SQL AI Agent for BI dashboards using fine-tuned LLMs, a React/Tailwind frontend, and a Django/Docker backend, delivering sub-5s queries deployed across multiple enterprise clients. The product became a core revenue-generating offering and reduced query turnaround time by 90%.

SDE Intern — Atomic Loops

April 2023 — December 2024

Engineered backend REST APIs using Django REST Framework for diverse products, facilitating data generation and collection for training ML models.
Implemented a real-time chat API using WebRTC for a client's website.
Designed and implemented an app for Yoga pose detection using Flutter.

ML Intern — Nishikawa Communications Pvt. Ltd.

August 2022 — April 2023

Developed an image straightener tool using OpenCV and TesseractOCR that automatically corrects tilted or flipped images, delivering results in 1–2 seconds.
Built a Japanese character classifier using transfer learning to recognize Hiragana and Katakana characters; containerized using Docker for easy deployment.
Increased efficiency by writing custom Python scripts for data analysis from the Google My Business API using pandas, and updated older scripts to match the newer API version.

Featured Projects

Tensorax

A standalone tensor computation library built from scratch with custom C++/CUDA kernels and Python bindings — no NumPy, PyTorch, or cuBLAS. Implements full automatic differentiation with a PyTorch-like API, neural network layers, optimizers, and loss functions. Custom CUDA matmul kernels (tiling, shared memory, memory coalescing) achieve up to 2.3× speedup over NumPy. Published on PyPI.

TinyGPT

A 95M-parameter LLM pretrained from scratch in PyTorch on a single RTX 3070 Ti (8 GB VRAM), using mixed-precision (AMP) and gradient accumulation to avoid OOM. Supervised fine-tuned on the Alpaca dataset for instruction-following. Deployed via Streamlit with dynamic model switching; weights on Hugging Face and live at tinygpt.streamlit.app.

AgentFlow

A multi-agent orchestration platform using LangGraph for intent classification, planning, and automated tool execution. Scalable workflow engine with custom scheduling (cron / interval), JWT authentication, and Docker-based deployment.

Split.ai

A fault-tolerant receipt parsing engine with a multi-LLM fallback strategy (Gemini → Llama 3), achieving 99% extraction reliability. Uses a graph-based debt simplification algorithm and Celery/RabbitMQ for asynchronous OCR processing.

QuillGPT

A GPT decoder block with self-attention and multi-headed attention implemented in PyTorch. Two pre-trained models — Shakespearean GPT and Harpoon GPT — with a Streamlit playground, FastAPI microservice, Docker containerization, and a custom tokenizer with portable serialization.

LoomRAG: Multimodal RAG

A multimodal retrieval-augmented generation system integrating OpenAI CLIP and Whisper for cross-modal retrieval across text, images, and audio, responding in 2–3 seconds. Streamlit interface supports annotation, custom datasets, CLIP fine-tuning, URL-based image indexing, and FAISS similarity search.

ParkVision

AI-driven parking analytics and customer segmentation dashboard. Fine-tuned YOLOv8 (0.968 mAP) and EasyOCR (85% accuracy) for vehicle plate recognition with IDFY API integration. Django REST Framework + FastAPI backend, React PWA frontend, PostgreSQL, Celery, Redis, Docker. GitHub.

DermaCare.ai

AI-driven portal for early detection of skin diseases in rural areas, with dermatologist consultation. Built a TensorFlow CV model deployed via FastAPI with S3 image uploads, plus offline TFLite inference in a Flutter mobile app. Web frontend in React, backend in Django REST Framework. Built end-to-end in a 24-hour hackathon.

M2ConneX

Alumni networking platform for MMCOE. Django REST Framework backend, React + Tailwind frontend, Docker-containerized. Real-time NSFW filtering via image classification and profanity detection (95% accuracy, 6–7s response). Custom Scikit-learn recommendation system for connections, posts, and jobs.

Sonnet

A chatbot that suggests songs from user input using LangChain and a FAISS vectorstore for lyrics similarity search, with 2–3s responses.

DevFinder

A social platform for college students and developers with AI-powered recommendations and real-time messaging. Flutter app, Django REST Framework backend, React frontend, TensorFlow recommendations, AWS deployment. GitHub.

OpenNN

A from-scratch Python neural network training library similar to TensorFlow 2, implementing layers, activations, losses, and optimizers. Install via pip install open-nn-python.

Blog Posts

Pre-training a 95M Parameter LLM on a Consumer GPU — walkthrough of pretraining TinyGPT on a single RTX 3070 Ti using AMP and gradient accumulation, plus SFT on Alpaca.
Writing CUDA Kernels from Scratch: A Beginner's Guide — the story behind Tensorax: custom CUDA matmul kernels, tiling, shared memory, and memory coalescing, with benchmarks.
Llama 4: The Natively Multimodal AI — what makes Llama 4 natively multimodal and how it differs from prior multimodal stacks.
Llama 2 Explained — transformer architecture, RMSNorm, grouped-query attention, and the modeling choices behind Llama 2.
Attention and the Versions of It: Layer by Layer — self-attention, multi-head, multi-query, grouped-query, and sliding-window attention compared layer by layer.
Attention: Explanation Is All You Need — an accessible explanation of the attention mechanism that powers modern LLMs.
From Words to Meaning: The Journey from Word Vectors to Learnable Embeddings — from one-hot encoding and Word2Vec to learnable embeddings inside modern neural networks.

Contact

Email: shrirangmahajan123@gmail.com. Find me on GitHub, LinkedIn, X, Hugging Face, and Kaggle.