Reinforcement Learning: Cart-Pole Balancing
Built RL system from scratch (Python/NumPy) using Monte Carlo Policy Gradient. Implemented custom physics sim, neural network, and training loop. Agent learned to balance inverted pole through pure trial and error.
Tech: Python, NumPy | Result: Stable convergence without frameworks
Real-Time Dynamic Hand Gesture Recognition
Built production-viable gesture recognition pipeline using TCN for temporal modeling. Achieved ~90% accuracy on dynamic gestures with real-time latency. Integrated MediaPipe for RGB input.
Tech: PyTorch, MediaPipe, NVIDIA TAO | Demo Video
WaveGAN: Generative Audio Synthesis
Implemented WaveGAN architecture in TensorFlow for raw audio waveform generation. Adapted DCGAN-style blocks for 1D time-domain signals.
Tech: TensorFlow, NumPy, Librosa
Age & Gender Estimation from CCTV
Solved Bosch challenge: demographic estimation from low-res CCTV footage. Simplified pipeline (removed face detection/super-resolution) achieved ~82% gender accuracy, ~66% age accuracy.
Tech: YOLOv5, ByteTrack | Demo
Multi-Agent RL & Game Theory
Mentored research project studying cooperation emergence in multi-agent environments. Implemented Q-learning agents interacting with fixed-strategy agents using evolutionary game theory principles.
Context: Semester project, Brain and Cognitive Society, IIT Kanpur