Building Autonomous AI Agents & LLM-Powered Apps |  LangChain • RAG • Local LLM Deployment | IBM Certified |

github

Abdul Moiz  • Peerlist

Technologies: Python, Java, TensorFlow, Pandas, NumPy, scikit-learn, pynput, TCP/IP Sockets, BizHawk Emulator, Joblib, Git LFS • Architected dual-language full-stack deep learning pipeline (Python + Java) with real-time TCP socket communication (60 FPS, <10ms latency) to BizHawk SNES emulator, capturing self-generated gameplay via streaming CSV architecture with pynput keyboard listener, implementing intelligent frame deduplication (20-40% redundancy reduction), and enabling competitive dual-bot gameplay on separate ports (9999/10000). • Engineered comprehensive training data collection and processing system capturing human gameplay recordings across 12 playable characters, transforming raw game state into 138-dimensional temporal vectors (6 frames × 23 features including derived position/health deltas), implementing 2-stage class balancing (2:1 sampling ratio + adaptive undersampling to 55% threshold) to address 80:20 idle-vs-active imbalance. • Trained 12 character-specific neural networks from self-generated datasets with multi-label sigmoid activation for independent 10-button prediction, implementing weighted loss functions (4x emphasis on button presses), low learning rate Adam optimization (1e-4), strategic dropout layers (0.3/0.2), and ModelCheckpoint callbacks achieving robust performance across diverse fighting game states while preventing overfitting. • Deployed real-time autonomous game AI integrating trained models with sub-frame inference (<10ms), implementing directional conflict resolution logic (UP/DOWN, LEFT/RIGHT opposition filtering), probability thresholding (0.5%), and frame-by-frame command generation, with unified GameState/Player/Buttons/Command abstraction layer enabling seamless emulator control and competitive multi-agent scenarios.