I researched and developed deep reinforcement learning (RL) models for options trading, implementing buy/sell/hold strategies for call and put options. The model incorporated variable capital management to allow multiple active positions simultaneously, with real-time tracking of P&L for each position and the overall portfolio. Within the deep RL model, I tested various architectures, including LSTM, GRU, and RNN, along with different reward functions for specific actions. Additionally, I integrated risk management techniques, such as variable capital allocation, take profit (TP), and trailing stop loss (SL), into the trading model.Beyond research and model development, I deployed the entire system on Azure using Azure Databricks. I created a complete data pre-processing and cleaning pipeline for tick-by-tick data from Upstox Websockets, which fed into the model forecasting. I then developed a pipeline for automated order execution based on the forecasted results, using Upstox APIs integrated with Azure Functions for real-time execution.

As a Quantitative Developer at Techsaq, I focused on cleaning and processing multi-time frame data for Bitcoin futures trading. I conducted research on statistical modeling and probability-based models, incorporating capital allocation strategies to manage multiple active positions with dynamic position sizing and risk management techniques. I tested and implemented models using various leverage levels, optimizing them to minimize liquidation risk through dynamic adjustments to capital exposure and leverage. Additionally, I developed backtesting dashboard templates for visualizing key performance metrics and trade patterns across quantitative strategies. I was instrumental in generating detailed PnL reports and tracking portfolio performance, factoring in transaction fees and platform charges for both long and short positions. I engineered signal generation modules leveraging technical and statistical indicators, and automated real-time trading execution on AWS using the CoinDCX API, optimizing trade execution across multiple time frames.

Developed a novel surveillance model using ConvLSTM and ResNet architecture to detect real-world violence in diverse scenarios. Deployed on the UCF-Crime dataset, capturing abnormal, illegal, and violent actions from public cameras, the model achieved 81.71% accuracy. Overcame challenges posed by varied lighting conditions and camera perspectives, demonstrating adaptability to complex surveillance settings.

During this internship at GrAI Robotics Private Limited, I worked on a project to detect and predict diseases in greenhouse plants using the input video feed. To achieve this, I built a ResNet-9-based model for classification, trained it on 76K RGB images of healthy and diseased crop leaves (covering 24 different classes of diseases) and achieved an accuracy of 89% on test images. The YOLOv5 algorithm was used to extract leaf images from each frame in the video or live stream, and images with similar properties were removed before predictions were made on the remaining images. The results were then integrated using the Streamlit and Flask framework and displayed on a website along with images of the leaves and their respective names and diseases.

During my internship at IIT Kharagpur and IIM Ranchi, I had the opportunity to work on exciting research projects under the guidance of Prof. J. Maiti at IIT Kharagpur and Prof. Sobhan Sarkar at IIM Ranchi. My work involved exploring and applying various optimization techniques related to the optimization of convex functions and quadratic programming. I also developed two kernel-free novel separation margins for binary classification, which were later extended to support vector machines (SVMs) resulting in the creation of two new SVM-based classification models, namely Q-SVM and Radius Bound Q-SVM (RBQ-SVM). I am proud to have successfully published two research papers on these models, one at the IEEE Conference on Data Analytics for Business and Industry in 2021 and the other currently under review in the Computers and Industrial Engineering journal.