Projects &
Skills

Featured Projects

• Audio Anomaly Detection System
  End-to-End Pipeline | CNN, Cloud Infra, Saliency Maps, Dash
  Fully automated audio monitoring system that collects microphone data, converts audio to spectrograms, and classifies anomaly types using a trained CNN. Built-in model interpretability via smoothed saliency maps, and a hosted Dash app for real-time inspection and playback of flagged events. Subscribed users are notified via automated email alerts, enabling hands-off anomaly tracking in production settings.
• Retrosynthesis Prediction
  Seq2Seq Modeling | Transformer, ReactionT5, Fine-Tuning
  Fine-tuned a domain-specific variant of Google's T5 transformer on retrosynthesis data to predict plausible chemical reactants from product SMILES. The model was trained using Seq2Seq learning with beam search and customized hyperparameters, treating molecules as a chemical language.
• Deep Reinforcement Learning for FPS
  Combat Strategy Learning | Dueling DQN, VizDoom, Reward Shaping, Depth-Aware Q-Learning
  Trained a Dueling DQN agent in a multi-bot Doom environment using grayscale + depth inputs. Designed a custom reward function to balance movement, accuracy, and survival, preventing local optima and encouraging strategic play. Used EMA-target networks, replay buffers, and Q-value analysis to refine learning.
• Ticketing System
  QR Code Infrastructure | Flask, ReportLab, SQLite, Nginx
  Designed and deployed a full-stack ticketing system with secure QR generation, HTTPS validation, and browser-based live scanning. Includes offline fallback, modular event config, and real-world use in high-volume entry management.
• Industrial Configuration Recommender
  Multi-Label Classification | Random Forest, Neural Networks, Postprocessing Pipelines
  Built a scalable recommendation system for medical device configurations. Optimized preprocessing and postprocessing to handle sparse high-dimensional data for reliable, explainable, and fast predictions in real-world industrial settings.
• Multi-Focus Occlusion Removal
  Deep Learning Architecture | CNN, SE Blocks, Upsampling, Custom MFOR Model
  Developed a deep CNN-based occlusion removal model (MFOR) using multi-focus image stacks. Included attention-enhanced skip connections and squeeze-and-excitation blocks to reconstruct occluded scenes with high PSNR. Managed 200k+ images and trained under hardware constraints for deployment in search-and-rescue imaging scenarios.

Test Scores

• TOEFL: 110/120
• Reading: 29/30 | Listening: 29/30 | Speaking: 26/30 | Writing: 26/30

• GRE:
• Quantitative Reasoning: 162/170
• Verbal Reasoning: 158/170
• Analytical Writing: 4.0/6.0