Autonomy software for a Mars regolith collection rover, connecting perception and state estimation with navigation and mining coordination.
Designed a ROS 2 architecture on NVIDIA Jetson, with LiDAR, IMU, and visual-odometry fusion. The navigation work combined SLAM, hierarchical A*/RRT planning, terrain classification, and local obstacle avoidance.
A state-of-charge estimator for CalSol’s solar vehicle, built to support energy management under changing driving conditions.
Developed an estimator using extended Kalman filtering and machine-learning techniques to predict battery performance and inform the vehicle’s real-time energy-management strategy.
For a complete background and work history, view full résumé (opens in a new tab)
Built BearTensor, a custom PyTorch-like tensor wrapper that tracks its operational history to dynamically build a Directed Acyclic Graph (DAG) during the forward pass.
Implemented Kahn’s algorithm for topological sorting to execute iterative backward-pass backpropagation without recursive stack overflow. Implemented gradient descent algorithms from scratch, including SGD, Momentum (tracking velocity matrices), and Adam (using β₁ and β₂ for adaptive moment estimation) to train a neural network predicting wine quality.
Implemented deep learning architectures from scratch for computer vision and sequence modeling, including fine-tuning across genomics and audio.
Built ResNet-18 utilizing residual learning blocks where g(x) = f(x) − x with stride downsampling and 1×1 convolutions for skip connections. Implemented a sequence-to-sequence Transformer with Scaled Dot-Product Attention, Multi-Head Attention, upper-triangular causal masking, and sinusoidal positional encodings. Fine-tuned pre-trained DNABERT-6 on genetic sequences using k-mer tokenization, and converted UrbanSound8k audio waveforms into spectrograms using STFT to fine-tune a ConvNeXt classification model.
Recreated the Chatbot Arena (LMArena) leaderboard by calculating relative model strengths from pairwise battle data.
Modeled the probability of Model A beating Model B as P(A beats B) = σ(S_A − S_B) via Logistic Regression without an intercept where features are +1 and −1. Resampled battle data iteratively using bootstrapping to generate 95% confidence intervals. Used TF-IDF vectorization to identify style hacking (verbosity and markdown formatting), penalizing overly verbose responses to produce a length-controlled leaderboard.
Designed a secure file storage and sharing client in Golang resilient against an adversarial, untrusted Datastore server.
Handled confidentiality using symmetric encryption (AES-CTR) and integrity using HMAC-SHA256. Secured user passwords with Argon2 memory-hard key derivation. Implemented RSA public-key cryptography to securely share file access keys, and designed key tree structures for efficient user revocation without re-encrypting large files.
Implemented search, adversarial agents, reinforcement learning, and probabilistic state tracking in stochastic, partially observable environments.
Implemented graph search (DFS, BFS, UCS) and A* search with custom admissible heuristics. Built Minimax and Expectimax adversarial agents with Alpha-Beta pruning. Implemented Value Iteration for MDPs, model-free Q-Learning with ε-greedy exploration, and Approximate Q-Learning using linear feature extractors. Tracked invisible moving ghosts from noisy sonar readings using exact Hidden Markov Model Forward inference and Particle Filtering with likelihood resampling.
Exploited vulnerabilities in compiled x86 C programs and implemented mitigation bypass techniques.
Exploited vulnerable C programs by overflowing stack buffers to overwrite the saved return address (EIP/RIP). Authored custom assembly shellcode. Bypassed ASLR using NOP sleds and information leaks, and bypassed Non-Executable (NX) stacks by chaining Return-Oriented Programming (ROP) gadgets to spawn a shell.
Simulated multi-body planetary orbits, orbital mechanics, and rocket launches using differential equation solvers in Python.
Solved systems of differential equations to model celestial orbits, gravitational interactions, and rocket launch ascent trajectories.
A teleoperation system for a 6-DOF robotic arm using ROS and Python, enabling real-time control and manipulation.
Developed a teleoperation system for a 6-DOF robotic arm using ROS and Python, enabling real-time joint-space and Cartesian manipulation with keyboard and mouse controls.
An end-to-end machine learning pipeline to identify faulty commits in GitHub repositories to improve code quality and streamline reviews.
Extracted features such as commit code changes and author metadata to train a classifier to predict whether commits are faulty or non-faulty.
Designed and optimized convolutional neural networks (CNNs) and trained models on the CIFAR-10 dataset using PyTorch.
Implemented key neural network components including vectorized backpropagation, batch normalization, and dropout, building hands-on deep learning workflows.
Designed and trained an LSTM-based neural network to generate realistic news headlines and a Transformer model to summarize lengthy news articles.
Trained an autoregressive LSTM network for sequential headline generation and implemented a Transformer architecture to effectively summarize news articles.