ABHRANIL DAS

Interested in and learning Data Science, Machine Learning and Artificial Intelligence, especially Computer Vision and Natural Language Processing. Skilled in Python programming and libraries like Numpy, Pandas, Scikit-learn, Pytorch, Tensorflow and NLTK

Currently exploring Generative AI with strong emphasis on LLMs.

LLM Prompt Recovery
An automated pipeline that uses LLMs to recover prompts from AI-generated texts using prompt engineering

• Used the Gemma 2-2B-Instruct model to recover prompts from AI-generated text
• Created a custom fine-tuning dataset from the Wikipedia Movie Plots dataset, consisting of movie plots rewritten using a wide range of prompts
• Fine-tuned the model using PEFT (QLoRA), achieving a ROUGE-L score of 0.37 and a Sharpened Cosine Score (SCS) of 0.59

Anomaly Detection in Surveillance Footages
An automated ML pipeline for the detection of violence, theft, and other anomalies in surveillance footages

• Built an end-to-end CNN + LSTM neural network model for detection of anomalies in surveillance footages using Tensorflow, and deployed using Gradio
• Used pre-trained VGG16 network for extracting spatial features from video frames and LSTM network for modelling the temporal relationship between frames
• Trained the model on clips from the Hockey Fight Detection Dataset, achieving an accuracy of 94% in anomaly detection on the validation dataset
• Extended the model for precise localization of anomalies in longer videos, with scope of extension to real-time monitoring and surveillance

CaptionCraft: Leveraging Transformers for Image Captioning
Enhancing visual storytelling by using transformers to generate intelligent and detailed image captions

• Created an image captioning application using Pytorch and Transformers (HuggingFace)
• Utilized pre-trained Vision Transformer (ViT) to extract high-quality features from images, and GPT-2 decoder to generate comprehensive and contextually-rich captions
• Fine-tuned the model end-to-end on the Flickr 8K Dataset, achieving a ROUGE-L score of 0.28 on the validation split

Anirban Dasgupta, Abhranil Das, Parishmita Deka, Soham Das, "Smart Embedded Systems - Advances
and Applications", Smart Cabin for Office using Embedded Systems and Sensors, CRC Press, Taylor and
Francis Group, pp.360 [2023] [Link]

• Notebooks Expert, Kaggle: Best-ever rank of 3053 among more than 19.6 M users on the platform
• JEE Advanced 2020: Secured AIR 1871 among 1,50,000 candidates appearing for the test
• KVPY SX 2019: Obtained KVPY fellowship by securing an AIR of 303