| Focus Area | Machine Learning & Foundations | Generative AI & LLMs |
|---|---|---|
| Architectures | SVMs, XGBoost, ResNets, Dense-connections, Residual Bottlenecks, ConvNets. | Hypernetworks, Diffusion, Transformer Blocks, MoE (Mixture of Experts). |
| Optimization | Init Schemes, Triton Kernels, Adam. | RLHF, DPO, GRPO. |
| Fine-Tuning | Hyperparameter Opt (Bayesian), Transfer Learning, Domain Adaptation. | PEFT: LoRA, QLoRA, DoRA. |
| Scaling & Infra | Distributed Training (FSDP, DeepSpeed ZeRO 1-3). | Inference: FlashAttention-3, vLLM, Speculative Decoding. |
| Model Compression | Pruning, Knowledge Distillation. | Quantization: 4-bit/8-bit (bitsandbytes, GGUF). |
| Theory & Math | Bias-Variance Tradeoff, Convergence Analysis. | Scaling Laws, Tokenization. |
| Data & Eval | Feature Engineering, PCA, AUC-ROC. | RAG, Synthetic Data, Reward Modeling. |
🎨 MMDiT-PyTorch | Multi-Modal Diffusion Transformer
A high-performance PyTorch implementation of the Multi-Modal Diffusion Transformer (MM-DiT) architecture, as featured in Stable Diffusion 3.
- Core Innovation: Implemented the "Symmetric Attention" mechanism where text and image modalities maintain their own learned embeddings while inter-modality information flows through a shared attention block.
- Engineering: Optimized for memory efficiency using scaled dot-product attention and modular block design to support varying sequence lengths across modalities.
- Theory: Integrated flow-matching objectives and time-conditioned modulation layers to achieve superior alignment between cross-modal representations.
⛓️ mHC-PyTorch | Manifold-Constrained Hyper-Connections
An implementation of the mHC framework, designed to stabilize and scale widened residual streams in deep foundational models.
- Core Innovation: Developed a projection-based connectivity layer that constrains residual mixing matrices to the Birkhoff Polytope (doubly stochastic manifold).
- Engineering: Integrated the Sinkhorn-Knopp algorithm for differentiable manifold projection and implemented kernel fusion techniques to minimize the memory access overhead caused by widened residual streams.
- Theory: Leveraged the closure property of doubly stochastic matrices under multiplication to guarantee stability.
- Languages:
Python,SQL - Frameworks:
PyTorch,TensorFlow,Transformers,Diffusers,Scikit-Learn,OpenAI,LangChain,LangGraph,LangFuse,LlamaIndex - Data Processing & Visualization:
NumPy,Pandas,Matplotlib, Seaborn - Databases:
MySQL,MongoDB,PostgreSQL - Vector Stores:
Azure AI Search,Qdrant,Weaviate,Vertex AI Vector Store,Pinecone - Cloud & DevOps Tools:
Docker,Kubernetes,Microsoft Azure,Amazon Web Services (AWS),Google Cloud Platform (GCP),Lambda Labs Cloud,Runpod,Minio,S3 - Machine Learning & Deep Learning:
Linear Regression,SVM,KNN,CNN,Transformer,GAN,Autoencoders,Reinforcement Learning,Large Language Models (LLMs),Diffusion,MLP - Applied Machine Learning:
Retrieval-Augmented Generation (RAG),Encoder-Decoder Modeling and Fine-Tuning,Contrastive Learning,Flash Attention,LoRA,QLoRA,DreamBooth,Hugging Face Ecosystem - MLOps:
FastAPI,Flask,ONNX,Weights&Biases,TensorBoard,MLFlow,KServe,TFServing,TorchServe,Llama.cpp,vLLM,Ollama,Nvidia Triton Inference Service,DeepSpeed,TensorRT,BitsAndBytes
- Email: nggkenny@gmail.com