Overview

I’m a mathematician with a strong interest in applied data science and its interdisciplinary impact. My background enables me to approach complex data-driven problems in a diverse array of fields like computational imaging/signal processing (PhD focus), customer behavior modeling, machine learning, DevOps, and financial risk analysis. I am very passionate about developing data-centric solutions that advance both technology and sustainability. I have outlined my relevant data scientific skills, background and project experiences below.

Skills

Machine Learning and AI: Supervised/Unsupervised learning, Forests, (Deep, Conv) Neural networks, XGBoost, etc.
Data Science: Regressions, Predictive modelling, Topological Data Analysis, Statistical/Stochastic Learning, etc.
Programming Languages: Python, R, SQL, Julia, C++.
Tools: PyTorch, Tensorflow, Scikit-learn, Numpy, Langchain, BI/Tableau.
Certificates: Data Analytics, Adv. Data Analytics (Google), Adv. Learning Algorithms (DeepLearning.ai), AWS.

Data Projects Summary

  • Predicting spatiotemporal dynamics of Chronic Wasting Disease (CWD) in KS

  • Link: Manuscript in preparation.
  • Goal: Develop a dynamic model to predict presence and spread of CWD in KS from highly unbalanced data.
  • Context: Project done as an inter-disciplinary collab. with the dept. of Veterinary Pathobiology and MUIDSI.
  • Deliverables:
    • Developed a (dynamic) spatiotemporal model capturing CWD movement in KS
    • Performed parameter-fine tuning through cross-validation with sampling data
    • Performance:
      • Max accuracy (fixed time, over space): 90.16%
      • Min accuracy (fixed time, over space): 85.11%
      • Overall accuracy (across space and time): 88.13%
    • Confirmed established trends, identified new migration pathways and suggested intervention measures.
    • Derived novel means to quantify the rate and direction of spread of disease
      • Developed ways to estimate rates and direction at different scales
    • Future work:
      • Aggregated KS soil covariate and land cover data and performed PCA to improve model performance
      • Esimated improved accuracy after covariate integration $\sim$ 93% [analysis underway]

  • Predicting passenger survival aboard Titanic - Kaggle competition

  • Link: Kaggle notebook
  • Goal: Predict passenger survival aboard the Titanic.
  • Context: Project done as a part of the Kaggle machine learning competition
  • Deliverables:
    • Placed in the top 1.3% globally with a bagged random forest model (with conditionally imputed trees)
    • Performed EDA and engineered nine new features, five of which were highly correlated with Survival
    • Evaluated various models with training accuracy as high as 83.3% under a 10-fold cross validation
    • Champion model: Bagged random forest
      • Test accuracy: 81.334%
      • CV standard deviation: 0.06%

  • Predicting user churn with Waze data

  • Link: GitHub
  • Goal: Predict user churn behavior by analyzing data collected from Waze app.
  • Data Source: Waze app (via Google)
  • Context: Project done as a part of an advanced data analytics certification powered by Google.
  • Deliverables:
    • EDA findings that indicated churn trends based on distance driven.
    • Engineered features that improved predictive power of the champion model.
    • Built regression and machine learning models that predicted user churn behavior.
      • Champion model: XGB-classifier
        • Accuracy: 81%
        • Recall: 16.5%

  • Harvest the sun! - Optimizing solar practicality across mainland US

  • Link: Blog post
  • Goal: Rank states/zip based on the socio-economic feasibility of installing solar panels to the median house.
  • Data source: Project sunroof, Google.
  • Context: Project done as a part of a data analytics certification powered by Google.
  • Deliverables:
    • Engineered three indices that captured unique aspects of solar feasibility.
      • Impact index: Amount of CO2 offset per installation.
      • Economical index: Short term costs and govt. subsidies for installation.
      • Savings index: Long term savings, EB offset from installation.
    • Designed a Tableau dashboard to communicate findings.

  • Nucleation of market bubbles

  • Link: GitHub
  • Goal: Adapt physical nucleation theory (JMAK) to the financial sector to predict market bubbles
  • Data source: FRED economic data
  • Context: Project done as a part of undergraduate senior capstone project
  • Deliverables:
    • Recontextualized the Avrami-JMAK equations, from physical nucleation theory, to the financial setting.
    • Fitted model to 2007 housing data to identify scale-invariant indicators of formation/collapse of bubbles.
    • Model RMS error: 12% over prior bubble phases of stocks from selected sectors.

  • Topological Data Analysis

  • Link: Blog posts - Part 1, Part 2, Part 3
  • Goal: Build a topological machine learning pipeline that classifies shapes based on topological features.
  • Context: Project done as a part of doctoral comprehensive exams on topological data analysis.
  • Deliverables:
    • Implemented a topological ML pipeline that
      • Extracted underlying topological information from point clouds (persistence barcodes).
      • Trained a random forest classifier on the topological features.
    • Achieved considerable dimension reduction by reducing number of training features from $O(3N)$ to $O(1)$, while retaining competitive performance.
    • Out-of-bag accuracy scores:
      • Synthetic data: 100%
      • Real life data: 82.5% (Source: Princeton computer vision course)

  • Robust Subspace Recovery

  • Link: GitHub
  • Goal: Extract a smaller dimensional subspace that contains “enough” points of a partitioned point cloud $\mathcal{X}$
  • Context: Project done as a part of doctoral comprehensive exams on quiver representation theory
  • Deliverables:
    • Implemented an algorithm that extracted a smaller linear subspace that contains enough points of $\mathcal{X}$
    • Furthermore, the extraction was simultaneous in the following sense:
      • $\mathcal{X}$ is formed by concatenating various point clouds into a single matrix
      • The extracted subspace contains enough points of this concatenation
    • Helps mitigate cases where such a recovery is not possible given only a single factor of the concatenation.

ML and programming

  • Solving mazes and simulating cycles - a saga of genetic algorithm

  • Link: GitHub - Maze Solver, ODE Parameter Estimator. Blog post - coming soon

  • Goal: Implement agents guided by the genetic algorithm to

    • Solve a randomly generated maze
    • Estimate the parameters of a coupled system of Ordinary Differential Equations

  • Deliverables: Developed agents (with hereditary genes) that

    • Solved a randomly generated maze in $O(M*N^2)$ time
    • Estimated parameters of the Predator-Prey system within $5$ generations of $5$ agents each.
      • Here is a sample run:

  • Numeripy - python package

  • Link: PyPi, GitHub

  • Goal: Develop a python package containing numerical ODE solvers and matrix methods

  • Deliverables:

    • numeripy.ODE_solvers offers numerical ODE solvers that offer robust precision control and flexibility
    • numeripy.matrix_methods offers matrix methods aimed for use in numerical linear algebra tools
    • In addition, numeripy.latexit() generates latex formatted tables ready to be pasted into a LaTeX document