Education
Master of Statistics [M.Stat.]
Indian Statistical Institute, Kolkata
Aggregate Score - Ongoing
Bachelor of Statistics [B.Stat.] (Honors)
Indian Statistical Institute, Kolkata
Aggregate Score - 72%
AISSCE (Class XII)
Hem Sheela Model School, Durgapur
Aggregate Score - 94.2%
AISSE (Class X)
Hem Sheela Model School, Durgapur
Aggregate Score - 95%
Publications & Pre-Prints
Hyperbolic Fuzzy C-Means with Adaptive Weight-based Filtering for Efficient Clustering
Accepted at the ICVGIP, 2025 in Oral and Poster category
Experience
Intern · Intelligent Shopfloor
Intelligent Shopfloor
Developing an intelligent pipeline for the mechanical industry utilizing neuro-symbolic AI.
Research Intern · TU Darmstadt, Germany
Self Organizing Systems Lab
Worked on the development of deep learning systems and generative frameworks for in-silico SELEX simulation.
Organizing Member · ISI, Kolkata
Winter School on Deep Learning [WSDL]
Contributed to an intensive academic event aimed at exposing students to cutting-edge research in Deep Learning.
Member · ISI, Kolkata
ISI Maths Club
Promoted interest and understanding in mathematics beyond the formal classroom setting.
Projects
Posterior Sampling from a Network via Rectified Flow
May, 2025 - Present
This work addresses Bayesian parameter estimation in network models incorporating both graph topology and node-level covariates. We propose a novel framework where the prior is implicitly defined through a parameterized transport mapping from a base noise distribution to a feature-conditioned target. To navigate the intractable posterior, we introduce a family of sampling algorithms that seamlessly integrate the complex implicit prior with the network likelihood. Our primary contribution includes both an exact iterative sampling scheme and a highly scalable approximate variant that relies on a closed-form analytical surrogate to bypass expensive numerical integration steps. Theoretical justifications for the proposed inference mechanisms are provided, alongside extensions for sparse network regimes.
Investigating Tree of Thoughts Prompting for Mathematical Reasoning in Large Language Models
December, 2025 - February, 2026
In this short-term exploratory project, we investigate the effectiveness of the Tree of Thoughts (ToT) prompting framework in enabling Large Language Models (LLMs) to solve simple mathematical reasoning problems more reliably than standard direct prompting. The project studies BFS- and DFS-based search strategies where the LLM generates and evaluates intermediate reasoning steps (“thoughts”) while navigating a structured search tree. Experiments are conducted on tasks such as the Game of 24 and elementary linear equation solving, with emphasis on reasoning consistency, search-guided problem solving, and interpretability of intermediate thought processes.
Escaping Preference Collapse: Cycle Dynamics in Nash Learning from Human Feedback
May, 2025 - May, 2026
Investigated Nash Learning from Human Feedback (NLHF) as a two-player zero-sum game to resolve non-transitive Condorcet cycles in LLM alignment. Conducted a rigorous dynamical-systems analysis, deriving a universal constant of motion for replicator dynamics and establishing topological bounds on policy orbits. Proved a strict Stability Bifurcation governed by the τ/η ratio in discrete Nash mirror-descent updates, extending theoretical guarantees to stochastic settings and symmetric k-action cycles. Validated these stability regimes empirically across synthetic tournaments, parameterized neural networks, and a 124M-parameter GPT-2 initialization.
Intersection of deterministic curves
May, 2025 - Present
In this project, we investigate the asymptotic behavior of self-intersection integrals for deterministic curves to establish rigorous scaling bounds. We initially ground our analysis in polynomial traversals, leveraging the density of monomials within the space of continuously differentiable functions with compact support. However, extending this framework reveals critical analytical hurdles, specifically a discrepancy between the upper and lower bounding rates caused by inherent traversal pathologies. To address these challenges, we introduce and define a novel 'regularized class' of functions. This regularization successfully resolves the identified bounding discrepancies, allowing us to formally establish matching upper and lower bounds that share the exact same stable asymptotic convergence rates.
Analysis of Low-Pass and High-Pass Filters on Graph Features
May, 2025 - August, 2025
In this study, we investigate the effects of low-pass and high-pass filters on the features of heterophilic and homophilic graphs. Low-pass filters are commonly used in signal processing and graph analysis to emphasize the smooth, low-frequency components of the data, whereas high-pass filters highlight the high-frequency components, effectively suppressing the smooth variations. We apply these filters to datasets characterized by either homophilic or heterophilic structures to derive insights into the graph structure and how transformations like these might aid in tasks such as classification, anomaly detection, or community detection.
Antithetic Noise in Diffusion Models
March, 2026 - Present
In this project, motivated from the recent ICLR 2026 paper Antithetic Noise in Diffusion Models, by Jia et al, we tried to understand how the authors describe the effect of antithetic sampling in Diffusion Models, which supposedly is a technique well prevalent in the literature of Variance reduction and try to interpret the concepts defined in the more general sampling frameworks.
A Generalized Framework for Penalized Local Mode-Seeking: Unifying Mean Shift and Sparse Smoothing Splines
November, 2025 - December, 2025
In this project we tried to look at the mode-seeking algorithms, such as the classic Mean Shift, which are foundational tools for non-parametric clustering and density analysis. We propose a generalized umbrella framework for mode-seeking, formulated as a Penalized Local Functional Maximization problem. We demonstrate that existing mode-seeking algorithms are special cases of this framework. Furthermore, we introduce a novel instantiation of this framework: the Sparse Smoothing Spline Mean Shift (SSS-MS). By employing a dual-penalty structure—an L2 integral roughness penalty and an L1 Group Lasso complexity penalty—the SSS-MS adaptively selects its own functional complexity at every spatial location, ensuring convergence properties that theoretically strictly dominate or equal the performance of its standard generalizations.
Bridging the Gap in Constrained Online Convex Optimization via the Polyak-Łojasiewicz Condition
November, 2025 - May, 2026
In this project we tried extending the optimal bounds in Constrained Online Convex Optimization (COCO) to overparameterized ML models by utilizing the Polyak-Łojasiewicz (PL) condition. Proved that a novel drift-plus-penalty framework successfully bypasses standard bounds to achieve logarithmic static regret and cumulative constraint violations.
Resampling Methods for Variance Estimation in G-Computation: A Comparative Study under Selection Bias
May, 2025 - October, 2025
In this project, we investigated variance estimation for g-computation under complex selection bias by replacing traditional, analytically derived M-estimation frameworks with three resampling techniques: non-parametric bootstrap, m-out-of-n bootstrap, and delete-d jackknife. Large-scale Monte Carlo simulations proved that these finite-sample resampling methods match the asymptotic sandwich estimator's performance, achieving nominal 95% coverage and calibrated standard errors. This approach trades computational speed for significant improvements in implementation simplicity.
Statistical Modelling of Prostate Cancer Biomarkers: A Comparative Regression Analysis
November, 2025
In this project, we did a statistical analysis of prostate cancer data, aiming to model the Log Prostate Specific Antigen (lpsa) using various clinical covariates. Utilizing a dataset of 97 patients, we employed Multiple Linear Regression, Robust Regression, Lasso Regularization, and Generalized Additive Models (GAM). Our findings suggested that despite the presence of influential observations, linear structures dominate the data generating process, with Lasso Regression providing the optimal balance between bias and variance. The study highlighted the critical role of tumor volume and capsular penetration in predicting PSA (Prostate Specific Antigen) levels.
Investigating Virtual Evolution: SELEX Simulator with Generative Models
May, 2025 - October, 2025
In this project, we worked on the development of a deep learning system for in-silico SELEX simulation and prediction. Designed and implemented a cross-attention predictor to model round-wise DNA sequence enrichment, and currently building a generative framework using flow matching and neural ODEs to simulate the full enrichment trajectory. Responsibilities include constructing embedding pipelines, training conditional vector fields, and evaluating models for predictive accuracy and generative performance. The work integrates bioinformatics, generative modeling, and dynamic system learning.
Convex Clustering methodologies
January, 2025 - Jun, 2025
In this project, We ventured through and explored various techniques that are prevalent in the landscape of the Convex clustering literature and various methodologies that commonly are used by practitioners to deal with the sort of optimization problems these usually transform into, namely ADMM [Alternating Direction Method of Multipliers], and using these we were attempting to adapt to situations where clusters w.r.t. both the rows and columns are of substance, a.k.a. biclustering, hence, devising an appropriate objective function(s) coupled with appropriate optimization subroutines to reach meaningful clusters.
Various Regression methodologies
August, 2024 - May, 2025
In this project, We have explored and assessed various kinds of penalizations prevalent in the literature of regression and how they affect the effectiveness of the model in terms of its ability in tactfully handling certain commonly observable traits in real life data and otherwise like sparsity, Multicollinearity and Model Misspecification.
Various Bregman Divergences and clustering based on those divergences and their consistency and convergence rates
October, 2024 - February, 2025
In this project, We have explored various Bregman divergences such as the Kullback Leibler divergence, euclidean distance and so on, and studied how they affect the clustering ability of an algorithm and how they are able to identify hidden structures or patterns in data without any human intervention or supervision. We also tried to analyse their convergence rates under various optimization subroutines like Gradient descent, ADAM and so on and hence obtain some asymptotic bounds on these rates along with which we also explored various convex formulations of such objectives and tried to observe how they influence each other.
On Reinforcement Learning, Evolutionary algorithms
January, 2024 - September, 2024
This Project is mostly about surveying literature in literature of RL and Evolutionary Algorithms and how these disciplines have intertwined and gave rise to the techniques commonly used. In this project, I implemented algorithms like minimax, Q-learning, MCTS (Monte Carlo Tree Search) and a few other variants, especially the evolutionary variants of it including a few inspired from deep learning methodologies on a few simple gym environments like tic tac toe, blackjack, ludo and a few more atari environments.
On the performance of PLUG-IN Method and LEAVE-ONE-OUT Method in construction of an estimate of the intensity parameter
April, 2024
This class project mainly focused on a comparative study of the PLUG-IN and the LEAVE-ONE-OUT Methods which are used quite often in estimation of the intensity parameter where, we want to obtain an estimate from the sample at hand, starting with a consistent estimate of the density (given, it exists), we construct a point-wise local average estimate of the parameter by considering a neighborhood. These methods intervene to determine the optimal size of the moving window for which the MSE of the estimate is in a reasonable vicinity of the Global minima over all the windows.
Exploring Spectral Clustering, Kernel based methods for clustering like kernel power K-means and Dirichlet Process Mixture Models
August, 2023 - December, 2023
This project is primarily centered around studying algorithms like K-means, Power K-means, Kernel K-means, Gaussian Mixture Models, where while initializing we need to state the number of clusters, and developing it into one where it itself determines the required number of clusters using Dirichlet Process Mixture Model like approach for optimal Loss, detects a broader variety of cluster instead of only the 'convex' ones like in Spectral clustering like algorithms and is less sensitive to the hyperparameters.
Reading Project on Spectral Clustering and its variants
May, 2023 - September, 2023
This project is mainly focused on studying and reproducing the traditional Spectral Clustering Algorithm as was suggested by Andrew Y. Ng, Micheal I. Jordan and Yair Weiss in their paper published on 2001 and a few variants of it namely, the one stated in the paper 'Eigen Selection in Spectral Clustering: A theory-Guided Practice' by Xiao Han, Xin Tong and Yingying Fan (accepted on April 2021).
Body Performance Analysis
May, 2023
This project is all about exploratory data analysis using multiple linear regression, univariate quantile regression and simulations of multiple variables for prediction of missing values on body performance data from a Korean sports promotion foundation.
Reading Project on Quantile Regression and Its Applications
February, 2023
This project is centered around study and comparison of exposure groups on the distribution of an outcome.
Bivariate Analysis of Nitrogen Use Efficiency and Nitrogen Fertilizer Use per hectare of cropland for India and Cuba between 1991-2014
December, 2022
This project is an exploratory analysis of Nitrogen Use Efficiency and Nitrogen Fertilizer Use per hectare of cropland and a quest for whether it is possible to cut down on the amount nitrogenous fertilisers used in agriculture without affecting the yield. This analysis provides us an assessment of the Cuban system of Organopónicos, and a vision for moving Indian agriculture towards sustainability.
Exploratory Data Analysis on Life Expectancy and GDP per capita of Russia, 1951-2022
November, 2022
This project is a preliminary analysis of how the life expectancy (in years) and the GDP per capita (in international dollars, fixed 2017 prices, PPP based on 2017 ICP) has been varying over the past 7 decades in Russia.
Relevant Coursework
Causal Inference, Optimization Techniques, Metric Topology and Complex Analysis, Measure Theoretic Probability, Nonparametric and Sequential Methods, Large Sample Statistical Methods...
Click to expand and view the full academic list →
Interests & Skills
Spatial Transcritpomics, Causal Inference and Discovery, Biostatistics, Flow Matching, Large Language Models, Bayesian Techniques, Statistical Learning, Non-Parametric Clustering techniques, Diffusion Processes, Deep Learning, Convex and Non-Convex Optimization Techniques, Convex Clustering, Evolutionary Algorithms, Reinforcement Learning, Graph Learning.
Languages
Tools
Academic Achievements
- Indian Statistical InstituteQualified for direct admission to M.Stat.(Master of Statistics) program with a government funded stipend.
- Indian Statistical InstituteAdmitted for B.Stat.(Bachelor of Statistics) (Hons.) program with a government funded stipend.
- Joint Entrance ExaminationsRanked amongst top 0.1 percentile in Joint Entrance Examination - Mains amongst 1 Million candidates.
Languages
- English: Full professional proficiency
- Bengali: Native or bilingual proficiency
- Hindi: Native or bilingual proficiency
Hobbies
Music, Table Tennis, Coding