AlphaConnect‑4, Prof. Vineeth N Balasubramanian

• Inspired by deep mind’s AlphaGo, implemented competitive multi‑agent Reinforcement Learning on connect‑4 environment.

• Utilized a combination of Monte Carlo Tree Search (MCTS) for opponent modeling and Actor Critic for agent reinforcement (single agent and single opponent). Designed the connect‑4 game environment as well.

• Achieved impressive results by training the agent on low‑dimensional board games and successfully applied transfer learning techniques to enable the agent’s performance in higher‑dimensional environments, all with minimal additional training.

VICAP: VIdeo Captioning And Prediction, Prof. Aditya T Siripuram

• Implemented a vision‑language video captioning method utilizing convolutional encoder with a attention based decoder.

• Engineered a three‑step search algorithm, employing Optical Flow techniques, to predict missing frames within video sequences. Additionally, exploited conditional Generative Adversarial Networks (GANs) for further frame prediction accuracy.

• Currently expanding capabilities in predicting missing frames within videos by exploring self‑supervised learning.

Gyro Correction in IMU sensors, Prof. K Sri Rama Murthy & DRDO India (Defence Research and Development Organisation)

• Spearheaded the creation of a gyro correction model for IMU sensors to mitigate noise and axis misalignment issues.

• Leveraged diverse architectural approaches, including DB‑LSTM, LSTM with attention mechanisms, and Transformer Encoder coupled with Huber Loss, while conducting rigorous training on the EUROC dataset.

• Through hyperparameter optimization, achieved superior performance with attention‑based models (Transformers), surpassing the capabilities of existing Dilated CNN methods in this domain.

Explaining Adversarial Examples & Robustness, Prof. Aditya T Siripuram

• Visual Explanations: Employed variants of Grad‑CAM and GRAD‑FAM techniques to produce insightful visual explanations for adversarial samples. Analyzed the behaviors of Convolutional layers to enhance model interpretability and robustness.

• Frequency Domain Analysis: Conducted in‑depth research into the frequency domain analysis of adversarial examples employing Fourier transforms and filters for MSIST and CIFAR‑10 datasets.

• Complex‑Valued Neural Networks: Currently involved in ongoing research focused on explaining adversarial examples within a frequency and complex space using complex valued neural networks

Metaverse, Bosch Hackathon

• Implemented a Metaverse Persona using UneeQ framework, and hosted this to local host and added customization to the webpage.

• Integrated google dialogue flow backend through webhook URL. Fed custom intents to make the persona respond accordingly to the user.

• In addition, integrated person identification, gender detection, and emotion recognition to the persona.

Open Face, Self project

• Worked on One‑Shot (Few‑Shot) Learning Facial Recognition using Siamese Network (Embedding Learning) on AT&T faces dataset.

• Trained and analyzed the performance of Prototypical Networks and Relation networks. Deployed the models in the open‑vino framework

Digital Pencil, Inter IIT Tech meet

• Invented a device translates hand Gestures into digital Characters. Created labeled data set using pyGARL, Arduino pro micro, accelerometer.

• Trained the model using linear SVM, kernel SVM, and ANN. Results show that kernel SVM and ANN outperform linear SVM in accuracy.