Projects
.png&w=3840&q=90)
Sentence-Level Bias Detection
Detects sentence-level media bias in news by separating emotion-driven bias from framing-driven bias. Compares sentiment (tone) vs TF-IDF (framing) and uses RoBERTa + classical ML to flag sentences biased through framing even when tone is neutral.
AI Travel Agent
Full-stack travel planning app with a Gemini-powered chat agent. Collects preferences, searches flights and hotels (Booking.com), food and museums (TripAdvisor), and builds an itinerary; sessions and suggestions stored in SQLite.
CartPole DeepRL Benchmark Suite
Unified benchmark of deep RL algorithms on CartPole-v1: value-based (DQN), policy-gradient (REINFORCE, Actor–Critic, A2C), and PPO. Each method in its own module with training scripts; shared networks and plotting in PyTorch + Gymnasium.
Network Motif Detection
Python implementation of RAND-ESU for sampling connected induced subgraphs (motifs) in large networks. Includes ESA baseline with bias correction, parallel execution, and significance testing with edge-swap ensembles.
VAE-GAN-RNN-arithmetic
Image generation with VAEs and GANs (CelebA), plus sequence-to-sequence models for arithmetic: text-to-text, image-to-text, and text-to-image (MNIST). Includes latent interpolation and recurrent encoder–decoder experiments.
Surrogate-model-eval
Surrogate model evaluation and model comparison: surrogate_model, IPL, LCCV, and vertical model evaluator for pipeline experiments. Includes comparison scripts and example runners.
Data Mining (ADM)
Advanced Data Mining course project: Python-based data analysis and experimentation with a written report. Covers core data mining concepts and applied analysis.
Flappy RL Benchmarks
Procedural Flappy Bird environment with dynamic themes (day, night, hell, space) and RL agents: PPO, DQN, DDQN, SAC. Built in Python with pygame; supports manual play and agent play for evaluating adaptation across difficulty and visuals.
Pacman Rangers
Group project: Python3 Pacman Capture the Flag (UC Berkeley CS 188 style). Design agents for two-team CTF in a Pacman arena—defend your food, steal the opponent’s, use power capsules and limited observations. Custom team logic, tournaments, and MCTS/planning.
Terrain-Adaptive House Generator
Automated pipeline extending GDPC: analyzes a 100×100 terrain slice, selects the flattest build site, and procedurally generates a garden plus a rustic cottage with randomized decorations and terrain visualizations in Minecraft.