I hear and I forget, I see and I remember,
I do and I understand

• Developed an AI support system to automate partner support and provide advanced analytics on service requests.

• Implemented a conversational AI agent using LangChain and OpenAI API that handles common partner inquiries, reducing human agent workload by 35% and improving response time by 60%

• Engineered a state management system using LangGraph to maintain conversation context and handle complex multi-turn human-in-the-loop interactions with tool-based actions, reducing human agent workload by 35%

• Built a real-time analytics dashboard with interactive visualizations and topic modeling that identifies issue trends, emerging issue patterns beyond predefined categories, enabling data-driven improvements to partner services. 

• Ranked Top 3 in EY IC25 Data Challenge for identifying locations and predicting severity of Urban Heat Island effect.

• Achieved a r-squared score of 0.9479 by implementing a Boosted Tree model with Residual Ensemble approach.

• Identified critical UHI drivers from Sentinel-2 satellite data and spectral indices through feature importance analysis.

• Experimented with DOFA [arXiv:2403.15356], GNN and K-Means clustering to explore multiple data modalities.

• Created GIS based visualization tools that overlaid UHI predictions to guide city planners with heat mitigation strategies. 

• Developed a Vision Transformer (ViT) powered image-to-GPS prediction system, achieving a mean geodesic distance error of 461.46 miles (742.65 km).

• Engineered a novel two-phase architecture, eliminating the traditional dependency on pre-stored address sets for classification.

• Integrated Google Base ViT and Swin Large ViT as feature extractors, producing custom embeddings based on the SOTA GeoCLIP image-to-GPS model.

• Implemented a highly accurate GPS decoder achieving average localization error of 12.55 miles (20.2 km). 

• Optimized data preprocessing by implementing NVIDIA DALI pipeline, accelerating training by 10x on a 1.2M image dataset, significantly reducing training costs. 

• Implemented LSTM based Recurrent Neural Network to predict S&P500 index, beating the market by a percentage point in a buy-and-hold strategy in a trading simulation on the SPY ETF. 

• Incorporated a compositional approach, leveraging component stock OHLC data to construct over 3,000 features per observation.

• Experimented with Recurrent Neural Networks (RNN)  and SARIMAX models, achieving a mean absolute error of 105.1

• This approach lays the foundation for a robust and adaptable trading strategy.

• Developed an NLP-powered document analysis platform using supervised ML algorithms (Random Forest, XGBoost) to classify textual similarity with 92% accuracy.

• Built comprehensive data pipeline with custom web scrapers (BeautifulSoup) to create a 25K+ document training corpus.

• Implemented advanced text preprocessing using NLTK and spaCy (tokenization, NER, dependency parsing) reducing feature dimensionality by 63%.

• Engineered hybrid feature extraction combining TF-IDF, Word2Vec, and BERT embeddings to capture statistical and semantic text properties.

• Demonstrated scalability by evaluating performance across multiple reference documents with a structured framework, enabling precise identification of plagiarized content with cosine similarity scores below a predefined threshold of 0.8.

Developed a complete irrigation system to tackle 3 main issues:

1. Erratic rain causing a scarcity of water resources. 

2. Unregulated use of water, leading to water waste.

3. Pests such as rodents causing crop damage.

This project implemented humidity sensing, rain sensing as well as pest control with multiple sensors relaying this information on to a smartphone using Bluetooth.