Oct 2025 – Ongoing

• Built a multi-stage computer-vision pipeline chaining YOLOv8 detection, DeepSORT multi-object tracking, a physics-constrained finite state machine, and transfer-learned 3D ResNet-18 spatiotemporal classification with automated CVAT annotation, converting 34+ hours of raw construction-site footage into per-subactivity energy profiles for mobile electric excavator charging dispatch.
• Achieved 88.81% recognition accuracy across 5 activity classes, surpassing a prior 3-class benchmark of 86.7%; applied Bayesian regression for hidden-state power estimation and built automated tooling for class-balanced clip extraction.
• Benchmarking deterministic certainty-equivalent MPC against scenario-based stochastic MPC over a 24-hour dispatch window to produce real-time mobile charging schedules that curb on-site battery depletion, demand charges, and grid carbon impact.

Tools: Python, PyTorch, YOLOv8, 3D ResNet, DeepSORT, OpenCV, CVAT, MPC, Bayesian Regression, Pandas, SciPy

Apr 2026 – Jun 2026

• Implemented dynamic programming for the Door-Key grid problems, computing optimal action sequences across 7 known and randomized MiniGrid environments.
• Built a 3D motion planner using weighted A* on a 26-connected grid with a parametric slab segment–AABB collision engine and forward-greedy path smoothing, clearing 7 environments (including dynamic multi-goal scenes) within strict per-goal timing budgets.
• Solved infinite-horizon stochastic optimal control for a differential-drive robot tracking a lemniscate trajectory via Receding-Horizon Certainty Equivalent Control (CasADi/IPOPT) with slack-variable obstacle avoidance and Generalised Policy Iteration on a discretised error-state MDP.
• Homework: worked through the planning & control foundations — deterministic shortest-path and label-correcting search, Markov Decision Processes with value and policy iteration, LQR, and reinforcement learning.

Tools: Python, NumPy, CasADi, IPOPT, Matplotlib, Gym-MiniGrid

Apr 2026 – Jun 2026

• Built StayTuned, a multimodal attention-detection system fusing a visual pipeline (MediaPipe landmarks → EAR/MAR/gaze/head-pose features → Bidirectional LSTM on DAiSEE) with an audio pipeline (Whisper transcription → ~75 acoustic features → LR/RF with speaker-aware GroupKFold) to classify students/drivers as attentive vs. distracted in real time.
• Fused per-interval scores via grid-searched weighting (0.85 video / 0.15 audio), supporting offline batch and live webcam + microphone inference, validated against ground-truth spreadsheets.
• Deployed a companion ESP32 + SSD1306 OLED and piezo-buzzer hardware alert for real-time attention warnings.
• Homework: implemented sparse linear regression (ISTA/LASSO regularization paths, GWAS marker selection) and Poisson GLMs via maximum likelihood; compared MLP vs. Transformer models for multivariate time-series forecasting; and built Neural ODEs and Fourier Neural Operators for the 1D wave equation with super-resolution analysis.

Tools: Python, TensorFlow, Keras, PyTorch, MediaPipe, OpenCV, Whisper, Scikit-learn, Librosa, ESP32/Arduino

Jan 2026 – Mar 2026

• Researched noise-conditioned controllability and reliable random seeds in text-to-image diffusion models, building on the All Seeds Are Not Equal (ICLR 2025) NPNet framework.
• Built a modular, resume-safe seed-mining pipeline generating 62k+ images with Stable Diffusion across numeracy and spatial prompt grids, with multi-GPU seed sharding, atomic writes, and OOM-fallback batching.
• Containerized the workload with Docker (CUDA) and deployed single- and multi-GPU Kubernetes jobs with persistent storage for large-scale generation.
• Homework: implemented eigenface PCA, Regularized Discriminant Analysis, and Gaussian classifiers for face recognition, and trained MLPs with backpropagation on MNIST (ReLU vs. sigmoid, SGD, learning-rate studies).
• Homework: built AdaBoost with decision stumps (one-vs-all ensembles with margin analysis) and kernel SVMs (LibSVM) on MNIST.

Tools: Python, PyTorch, Hugging Face Diffusers, Stable Diffusion, Accelerate, Docker, Kubernetes, MATLAB, LibSVM

Jan 2026 – Mar 2026

• Implemented quaternion-based 3D orientation tracking from IMU data using Projected Gradient Descent on the SO(3) manifold with gyroscope-bias learning, then stitched calibrated camera frames into equirectangular panoramas.
• Built a full 2D SLAM pipeline: differential-drive odometry, SVD-based ICP scan matching, log-odds occupancy grid mapping with Bresenham raycasting, RGB-D texture mapping, and GTSAM factor-graph optimization with loop-closure detection.
• Developed a Visual-Inertial SLAM EKF on SE(3) fusing IMU prediction with stereo landmark updates — Shi-Tomasi + Lucas-Kanade feature tracking, stereo triangulation, a Mahalanobis innovation gate, and a numerically-verified left-perturbation SE(3) pose Jacobian with sparse joint pose-landmark covariance.
• Homework: derived the core estimation theory — Bayesian and Gaussian filtering, rigid-body motion and rotations on SO(3)/SE(3) with quaternions, and the Kalman, Extended Kalman, and particle filters.

Tools: Python, PyTorch, NumPy, SciPy, OpenCV, GTSAM

Jan 2026 – Mar 2026

• Built SolarAgent, an agentic LLM tool (LangGraph) pairing a natural-language front-end with a MILP optimization backend to recommend residential solar PV and battery sizing for real San Diego households.
• Formulated and solved a mixed-integer linear program (HiGHS) over an 8760-hour annual dispatch, jointly sizing panels and batteries from real hardware catalogs (Tesla Powerwall, Enphase, SolarEdge) under TOU/demand tariffs, roof-area limits, budgets, federal ITC, and EV charging.
• Also built a zero-cloud LLM + TF-IDF RAG advisor (Track A) for PV sizing, batch-analyzing 30 San Diego neighbourhoods with configurable local backends (Ollama/vLLM), and evaluated the agent against fixed baselines on Kubernetes.
• Presented a paper on RLHF / InstructGPT (Training Language Models to Follow Instructions with Human Feedback).

Tools: Python, LangGraph, HiGHS (MILP), Ollama, vLLM, LLaMA, RAG (TF-IDF), NumPy, Pandas, Kubernetes

Sept 2025 – Dec 2025

• Built an end-to-end statistical learning pipeline to predict national happiness scores by integrating multi-source global datasets (World Happiness Report, World Bank, CO₂ emissions, population data).
• Applied feature engineering and statistical validation, including factor analysis for maternal health indicators, VIF-based multicollinearity control, and Bonferroni correction for multiple hypothesis testing.
• Trained and evaluated regression models (Ridge, Random Forest, Gradient Boosting), achieving R² = 0.88 and identifying key predictors such as social support, GDP per capita, and life expectancy.
• Conducted rigorous exploratory data analysis and scaling comparisons (Standard, Robust, MinMax), extracting interpretable insights on economic, social, and environmental drivers of well-being.

Tools: Python, Overleaf(Latex)

Sept 2025 – Dec 2025

• Implemented Bayesian classifiers for cheetah vs. grass image segmentation using DCT features, achieving minimum probability-of-error decisions via histogram-based and Gaussian models.
• Modeled high-dimensional data with multivariate Gaussians (ML & MAP) and demonstrated the curse of dimensionality, reducing error from 5.52% (64D) to 3.11% (selected 8D features).
• Developed fully Bayesian predictive classifiers with informative and neutral priors, showing improved robustness on small datasets and convergence behavior as sample size increased.
• Trained Gaussian Mixture Models using Expectation-Maximization, analyzing initialization sensitivity and model complexity to identify an optimal bias–variance tradeoff (C = 8 components, ~4% error).

Tools: Matlab, Overleaf(Latex)

Sept 2025 – Dec 2025

• Designed an autonomous roadside mechanic that detects broken-down vehicles via blinking hazard lights and navigates to park safely behind them using multi-sensor fusion.
• Implemented a ROS2-based distributed system with a Jetson Nano (bird’s-eye vision) and Raspberry Pi (on-vehicle perception & control) communicating through custom publish-subscribe nodes.
• Trained and deployed a Roboflow/YOLO deep learning model (~90% accuracy) on an OAK-D camera to detect the rear of vehicles and compute angular offsets for PD-based navigation.
• Integrated LiDAR-based distance estimation and vision-based control to enable precise stopping behavior behind the disabled vehicle while reducing false positives using HSV filtering, temporal analysis, and FFT-based blink detection.

Tools: Python, Linux, ROS2, OpenCV, Path Planning, Sensor Integration, Embedded C, Computer Vision, Deep Learning, Roboflow, YOLO, Jetson Nano, Raspberry Pi, LIDAR, GPS, VESC, DC-DC converter

Jan 2023 – May 2023

• Modeled ANN, LSTM, and RNN to predict power loss at different states of charge and electrolyte flow rates.
• Improved accuracy of existing models by 40–70%.
• Optimized energy storage systems for cost and efficiency.

Tools: Python, TensorFlow, PyTorch

Mar 2023 – May 2023

• POSIX-compliant multithreaded server with shared memory IPC.
• Stateless client-server communication: register, request, response, unregister.
• Logging of all intermediate states and request counts.

Tools: C, POSIX threads, Shared Memory

Jan 2023 – Mar 2023

• POSIX program for 2D matrix operations using 1-to-1 pipe between controller & workers.
• Fork + multithreading for efficient row-wise calculation.
• Handled errors and signals robustly.

Tools: C, POSIX, Multithreading

Jan 2023 – May 2023

• Scheduler for BITS community to manage availability and meetings.
• Members register as staff/students using BITS email.
• Set personal availability and view availability of others.

Tools: Java, Intellij, MySQL

Jan 2023 – May 2023

• Plotted central and implicit finite difference scheme results for 4 wave propagation speeds using Richter wavelet.

Tools: Python, MATLAB

Jan 2023 – May 2023

• Implemented NPC converter to increase output voltage vs. boost converter.
• Implemented MMC converter to improve efficiency of wind generation.

Tools: MATLAB Simulink, Power Electronics Toolbox

Nov 2022 – Dec 2022

• Aggregated 30+ feature variables, preprocessed, and analyzed patient data.
• Baseline logistic regression improved with Decision Trees, Naïve Bayes, and SVM.
• Achieved >99% accuracy after hyperparameter tuning.

Tools: Python, Scikit-learn, Pandas, NumPy

Oct 2022 – Nov 2022

• Collected data from 30 smartwatch users.
• Preprocessed and trained model to predict wearer activity with high accuracy.
• Tested in real-world scenarios.

Tools: Python, Scikit-learn, Pandas

Nov 2022 – Dec 2022

• Implemented hash table-based spell checker.
• Corrects common errors: swap, insert, delete, replace characters.

Tools: C++

Oct 2022 – Dec 2022

• Developed ML model for wind power projections with superior accuracy vs. basic time series models.
• Used BCD classifier + SVR for non-stationarity and adaptability.
• 48-hour predictions outperform persistence model.

Tools: R

Nov 2022 – Dec 2022

• Designed half-adder using NAND, NOR, Inverter gates at transistor-level CMOS.
• Combined two half-adders into 1-bit full adder.
• Optimized propagation delay and power efficiency.

Tools: LTSpice

Aug 2022 – Dec 2022

• Implemented Piecewise Lagrange Linear and Quadratic approximations and interpolations.

Tools: MATLAB

Mar 2022 – Apr 2022

• Used game theory to predict optimal pit stops under dry, wet, and safety car conditions.
• Calculated Nash equilibria and payoffs for different tyre strategies.
• Applied model to real-life race scenarios.

Tools: Python, MATLAB

Jan 2021 – May 2021

• Neural network trained on bank queue data to predict waiting times (MAE: 2.68 minutes).
• Verified queue-specific networks using web app simulator.
• Generalizable to other industries.

Tools: Python, Keras, TensorFlow