| 🔬 Titan Architecture Custom DistilBERT variant with 40% fewer parameters and 79% SST2 accuracy. 🔗 View Project |
| 🤖 model2gguf Convert Hugging Face models to GGUF (int4/fp16) for Ollama-based deployment. 🔗 PyPI Package |
| 💡 RL API Fast PPO runner with <500ms latency using Flask; optimized for memory-constrained environments. 🔗 View Project |
| 🚀 CUDA-ML High-performance CUDA C++ kernels for ML — matrix ops, memory-efficient convolution, and PyTorch benchmarking. 🔗 View Project |
| 🧬 IEEE Research Brain Tumor Detection using InceptionResNetV2 (99.45% accuracy) with augmentation + recall-optimized loss. 🔗 IEEE Paper |
| 📱 Parkinson's FOG Detection Real-time LSTM classifier running on Raspberry Pi 4; 65% memory optimization from preprocessing pipeline. Private / Embedded Deployment |
- 🎨 Diffusion models for spatial-temporal alignment in T2I (CMU CV Lab)
- 🚀 CUDA warp-level optimization for memory-efficient convolution
- 🧬 Hybrid Quantum-Classical ML for tumor classification (PennyLane + PyTorch)
- 🧠 Fine-tuning LLMs + building custom RAG pipelines with QLoRA & LlamaIndex