Sri Lasya Siripurapu

Problem: Visually impaired users often lack immediate, descriptive context of their surroundings from images.

Solution: Architected a real-time assistive technology using a GPT-4 Vision model for its state-of-the-art scene understanding and an ESPnet TTS model for natural language narration, all packaged in a lightweight Streamlit app for accessibility.

Impact: Delivered audio captions with an average end-to-end latency of ~3.5 seconds, enabling users to receive near-instant feedback and navigate their environment more independently.

Problem: Standard CNNs can overfit on small datasets, leading to poor generalization and unreliable feature extraction for classification tasks.

Solution: Designed a novel supervised autoencoder that combines unsupervised feature learning (reconstruction) with supervised classification by integrating class labels into the bottleneck layer, forcing the model to learn a more robust and informative latent space.

Impact: Increased image classification accuracy by 38% over a baseline CNN, creating a more generalizable model valuable for domains like medical imaging where interpretability and reliability are critical.

Problem: Traditional financial models often fail to capture real market volatility due to their reliance on fixed assumptions.

Solution: Applied stochastic calculus to forecast stock prices, implementing Monte Carlo simulations to model thousands of potential outcomes and the Black-Scholes formula for precise European option pricing.

Impact: Created a robust tool for quantitative analysis that enables more realistic pricing and better hedging strategies, essential for modern risk analysis and portfolio management.

Problem: Building a foundational understanding of Neural Machine Translation (NMT) requires moving beyond pre-trained APIs to implement core architectures from scratch.

Solution: Engineered an English-to-Spanish translation model using a Bi-LSTM Seq2Seq architecture with an attention mechanism, chosen specifically to handle long-range dependencies in language.

Impact: Achieved a BLEU score of 0.23 on a 35k+ sample dataset, demonstrating a strong, foundational grasp of sequence modeling, attention mechanisms, and the principles of modern NLP systems.