Aryan Bibhuti

Room No. B-111, Patel Hall, Indian Institute of Technology Kharagpur, West Bengal, PIN - 721302, India
# aryanbibhuti402@gmail.com ï LinkedIn § GitHub
Education
Indian Institute of Technology, Kharagpur Kharagpur, India
Integrated Dual Degree (B.Tech+M.Tech) in Computer Science and Engineering CGPA: 9.22/10 2021-2026

Research Interests
Large Language Models, Natural Language Processing, Geometric Data Science, Deep Learning

Research Experience
Tr2 AIL Lab, IIT Kharagpur Jun 2024 – Present
Investigating World Models in LLMs | Advisor: Prof. Somak Aditya Kharagpur, India
• Analyzed Key Research on Reasoning and World Modeling in Language Models by Investigation on Rational

Speech Act (RSA), Elements of World Knowledge (EWoK), and Fermi Problems, to identify gaps and opportunities for
improving LLMs
• Perturbed GRICE and CICERO datasets to assess LLM performance and highlight Limitations in World Modeling

• Tested GPT-3.5, GPT-4o mini, and Microsoft Phi Language Models on perturbed datasets to gauge their

Pragmatic Reasoning capabilities using Robust Accuracies as the primary metric

• Presently focusing on Fine-Tuning Language Models using Direct Preference Optimization (DPO) to enhance

their World Modeling and Pragmatic Reasoning abilities.

HCI-BCI Lab, IIT Kharagpur May 2024 – Present
Brain Connectivity Analysis | Advisor: Prof. Debasis Samanta Kharagpur, India
• Worked on temporal and spectral aspects of connectivity on the motor-imagery dataset using mne python library,

resulting in a 15% improvement in classification accuracy

• Computed statistical metrics such as partial correlation coefficient (mean = 0.78), mutual information (mean = 0.35

bits), and canonical correlation (mean = 0.62)

• Current work includes thresholding the adjacency matrices, performing graph analysis, and implementing DL Models

Research Intern, University College London Dec 2023 – Apr 2024

Machine Learning | Advisor: Prof. Philip Treleaven London, UK
• Developed a Time - Series model for forecasting stock price movements using Long Short-Term Memory on LSE

data
• Implemented a RoBERTa based text-mining tool for discovery of hidden semantic structures in a text body

• Achieved a ROUGE-1 score of 0.259 to extract and summarize the contents of financial reports of company

Research Intern, University of Turku Dec 2023 – Feb 2024

Deep Learning | Advisor: Prof. Abdulhamit Subasi Turku, Finland
• Designed architectures of 2 and 3 layered CNNs to obtain accuracy of around 95% for classification of Brain MRI scans

into three different stages of Alzheimer’s - Very Mild, Mild, and Moderate.
• Achieved an accuracy of 96% by fine tuning VGG-16 architecture

• Combined Machine Learning models and ResNet-50 pre-trained architectures to enhance the above accuracy to more

than 96.5%

Key Projects
Multi-class Document Classifier Sep 2023 - Nov 2023
Github | Tensorflow
• Developed a multi-class image classification model for classifying grayscale documents from the RVL-CDIP dataset

• Boosted accuracy score by 15% by appending a Supervised Contrastive loss function to the ResNET module

• Improved acc by 29% with bootstrap aggregation of MobileViT, MobileNet models achieving 77% overall accuracy

Image Style Transfer Feb 2023

Github | Pytorch
• Implemented style transfer technique from the paper “Image Style Transfer Using Convolutional Neural Networks” to

generate using two images a new image that reflects the content of one but the artistic ”style” of the other
 • Used SqueezeNet to extract features and used formulated loss function to perform gradient descent on the input
Instance Segmentation Module Jan 2023
Github | Pytorch, Tkinter, OpenCV, Matplotlib
• Created a package for transforming images and analysing their effects on the predictions of an instance segmentor

• Employed MaskRCNN model pretrained on COCO dataset that outputs segmentation masks, confidence & boxes

• Wrote plot-visualization utilities to plot segmentation masks and bounding boxes, implemented the GUI in Tkinter

Early Pneumonia Detection Using X-Ray Lung Scans Jul 2022 - Aug 2022
Github | Keras
• Created a classifier model using 2-D Convolutional Neural Networks to detect early pneumonia from X-ray lung scans

• Achieved Recall score of 98% and Precision of 83% by fine tuning VGG-16 Pre trained Architecture

Machine Learning Nanodegree Capstone Project | Udacity Jun 2022 - Jul 2022
Github | Scikit-learn, Matplotlib
• Developed plagiarism detection model (Accuracy = 0.96) to identify whether two given texts belong to same source or

not
• Fine tuned RoBerta (Accuracy = 0.81) for identifying the legitimacy of natural calamity tweets

Relevant Coursework
Deep Learning: Stanford: CS229, CS231N, DeepLearning.AI: DL Specialization, Jovian: DL with PyTorch
Mathematics: Probability & Statistics, Linear Algebra, Advanced Calculus, Numerical & Complex Analysis
Computer Science: Machine Learning, Deep Learning, Reinforcement Learning, Computer Networks, Database
Management Systems, Operating Systems, Compilers, Systems Programming, Computer Organization & Architecture,
Formal Language & Automata Theory, Software Engineering, Algorithms- I & II, Programming & Data Structures

Technical Skills
Languages & Frameworks: Python, C, C++, gawk, LaTeX, Java, MATLAB, Django, Flask
Libraries: PyTorch, Tensorflow, Scikit-learn, NLTK, spaCy, HuggingFace, Matplotlib, Seaborn, Backtrader, OpenCV, PIL
Developer Tools: gcc, make, gdb, valgrind, gprof, grep, bash, Git, VS Code, Jupyter Notebook, MS Office

Competitions
Data Analytics Challenge - Data Science Summit’23 - BIT Mesra | Rank 1/538 Mar 2023
• Implemented techniques such as EDA, feature selection using pearson correlation and implemented different machine
learning algorithms such as XGBoost, AdaBoost etc and achieved an accuracy of 90.11%
Cascade Cup, IIT Guwahati | Rank 5/2000+ Jan 2023
• Implemented different machine learning models(SVM,XGBoost) to achieve a score of Weighted F1-score of 98.53 %
in predicting the age group of user on the basis of their tweets and other internet activities
Data Unchained - IEEE IIIT Delhi | Rank 4/150+ Oct 2022
• Stood 4th out of over 150 participating students in a 4 day tri-hackathon involving employee attrition rate prediction
(Auc Roc Score = 0.83), social media emotion recognition (Categorical-Accuracy = 0.39) and vehicle speed detection
using opencv (RMSE = 2.15)

Academic Achievements
• Secured All India Rank 408 (GE) in JEE Advanced, 2021 among more than 2,00,000 candidates across India
• Secured All India Rank 434 (GE) in JEE Mains, 2021 among more than 10,00,000 candidates across India
• Awarded Kishore Vaigyanik Protsahan Yojana (KVPY) Fellowship by DST, Government of India
• Awarded National Talent Search Examination (NTSE) Fellowship by the NCERT, Government of India
• State rank 35 in Regional Mathematics Olympiad (RMO) and qualified for Indian National Mathematical
Olympiad (INMO), which is the third round to select students from India for International Mathematics Olympiad
(IMO)