I’m a data scientist at Chubb in Philadelphia, building predictive models, machine learning tools, and agentic AI applications. My work spans insurance pricing, conversion and retention modeling, and tools that help teams move from data to decisions. I hold dual master’s degrees in Robotics and Computer Science from the University of Pennsylvania, where I worked on perception and localization for an autonomous racing go-kart. Beyond my day job, I built PeeqAI, a multimodal assistant for visually impaired users, and co-founded Philly AI Connect. I’m especially interested in how vision, language, and learning can come together to solve practical problems.
North America Data Science Modeling team, working on insurance pricing, predictive modeling, and internal AI applications.
A community bringing together Philadelphia’s AI and machine learning practitioners, researchers, and builders.
Oct 2021 - May 2023, Philadelphia, PA
Research in autonomous systems at xLAB, alongside teaching in applied machine learning and community coding education.
Dec 2021 - May 2023
Jan 2022 - Jul 2022
Oct 2021 - Jul 2022
Oct 2019 - Jun 2020, Chennai, India
Research in embedded systems, assistive technology, and robotics.
Dec 2019 - May 2020, Bengaluru, India
An IoT farming startup connecting sensors, cloud infrastructure, and mobile applications.
Wind energy engineering and embedded sensing applications.
Aug 2021 - May 2024 Dual M.S.E. in Robotics and Computer and Information ScienceGPA: 3.8 out of 4Taken Courses
| ||
2016 - Jun 2020 B.Tech. in Electrical and Electronics Engineering with a Minor in Computer ScienceGPA: 8.99 out of 10PublicationsTaken Courses
Extracurricular Activities
|
A hands-free assistant for visually impaired users, combining a custom vision-language model, wake-word speech recognition, and GPT-4 for real-time scene understanding. A context layer lets the assistant reason over accumulated environmental state.
Built perception and localization with PSPNet, GNSS/IMU fusion, and TensorRT on NVIDIA Jetson AGX. Led the team to first place at Autonomous evGrandPrix 2023 and presented at ICRA 2022.
Production GLM and XGBoost models for pricing, quote conversion, retention, and loss cost, alongside reusable modeling tools and customer lifetime value applications.
Agent workflows for structured extraction and natural-language analytics, helping teams explore portfolio data and automate modeling tasks.
A D&D rules and Dungeon Master companion with cited hybrid retrieval, campaign memory, and dice tools. Combines a React interface with a FastAPI backend and support for local or hosted language models.
A real-time visual assistant that combines a phone video stream, CogVLM image understanding, GPT-4 reasoning, and spoken interaction. Keeps recent visual observations as context for answering questions.
Ongoing conversations between GPT and Google Gemini, with the option for a person to join the exchange.
A comparative study of Google Gemini Vision and CogVLM for vision-language tasks.
A Streamlit chatbot prototype using LlamaIndex to index documentation and retrieve relevant context for GPT-powered answers.
My Hugo and Toha portfolio, with professional projects, writing, a photo gallery, and interactive games and sandboxes. Uses a shared Vite build and Netlify functions for its AI-powered experiences.
Improved state-of-the-art EfficientPS model by replacing its instance head (Mask RCNN) with SOLOv2.
Benchmarking 9 ML algorithms for identifying hand motions from EEG recordings.
PyTorch implementation of proposed improvements to the ESRGAN image super-resolution model.
A deep learning approach to blind motion deblurring that estimates blur parameters for Wiener deconvolution, improving the legibility of blurred license plates. Includes an interactive demo on Hugging Face.
Robot localization using LSTM in HabitatAI simulator, employing particle filter as the embedded algorithmic prior.
Variational Autoencoder that learns from MNIST dataset and generates altered handwritten digits.
Indoor SLAM using an IMU and a LiDAR sensor of a humanoid called THOR.
Bottom-up implementation of YOLOv3 for Object Detection.
PyTorch implementation of Segmenting Objects by LOcation (SOLO) and Feature Pyramid Networks for the COCO dataset.
Recurrent Neural Network trained on Leo Tolstoy’s War and Peace to imitate his language style.
Detects spam emails with 98.7% accuracy using Support Vector Machines
Numpy implementation of a neural network for digit recognition
Robot orientation tracking using EKF and UKF on VICON motion capture system
An autonomous robot that scans its surroundings to locate and extinguish fires.
A Python implementation of the Baum-Welch algorithm for learning the parameters of a hidden Markov model.
Experiments with the ML Bias Bounty workflow: defining population groups, training group-specific predictors, and submitting improvements to a shared Pointer Decision List model.
A Python implementation of the multistage Canny edge-detection algorithm for identifying image boundaries.
Three algorithms for solving the knight’s tour problem, exploring how a knight can visit every square on a chessboard while managing time and space complexity.
An all-convolutional network trained on CIFAR-10 in PyTorch to study cosine annealing for learning-rate scheduling.
A convolutional neural network that colorizes grayscale images, with PyTorch training and inference code, a sample dataset, and color-temperature control.
Handwritten-digit classification experiments combining Gabor image features with support vector machines on MNIST.
A fork of UPenn GoKart’s documentation repository, with Sphinx and reStructuredText sources for the autonomous go-kart project.
A comparison of gradient descent, stochastic gradient descent, and Nesterov momentum for logistic regression on MNIST.
A lane-detection pipeline using camera calibration, perspective transforms, color and gradient thresholds, and polynomial fitting to estimate lane curvature and vehicle position.
Image blending with Gaussian and Laplacian pyramids, using multiscale reduce, expand, combine, and collapse operations.
Exploratory notebooks and data preparation for protein reverse engineering, including E. coli data.
Policy iteration for finding a robot trajectory from a start cell to a goal in a two-dimensional grid world.
Q-learning with the Double DQN technique for balancing a pole on the CartPole control task.
A Turbo C++ implementation of tic-tac-toe with single-player and multiplayer modes.
A Bayes filter that tracks the position of a robot moving through a two-dimensional grid world.
A simplified implementation of a vision transformer in PyTorch.
A multimodal visual question-answering model combining CNN image features, an LSTM question encoder, and stacked attention, with visualizations of the attention layers.
An Arduino-based wrist-worn prototype that monitors heart rate and motion from accelerometer and gyroscope readings to estimate sleep quality.
An initial repository for image-colorization work. No implementation has been published in this repository yet.
The source for my GitHub profile README, including project links and technical interests.
