(quotes intentional, impostor syndrome included at no extra cost)
I’m a burnt-out programmer living at the edge of generative modeling and decision science, where GPUs scream, losses explode, and convergence is a suggestion. These days I’m obsessed with making SLMs/LLMs scale without crying, pushing Diffusion Transformers (DiT) past their U-Net era, and applying Bayesian methods so my models can say “I’m not sure” with confidence.
- LLMs & SLMs: Designing robust pipelines for Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to align models with complex reasoning tasks.
- Bayesian Forecasting: Leveraging Gaussian Process Regression (GPR) for uncertainty-aware predictions in high-stakes environments like infrastructure management.
- Time-Series Hybridization: Combining classical VARIMA models with Deep Learning (MLP) for multi-variable forecasting.
- Diffusion Transformers (DiT): Investigating scalable transformer-based backbones for latent diffusion, moving beyond U-Net architectures for higher-fidelity generation.
- RAG Systems: Developing Multimodal RAG architectures capable of synthesizing insights from variant data sources, including structured tables, unstructured text, and visual assets, utilizing cross-modal embeddings for holistic retrieval.
- Post-Training & Alignment: Engineering iterative workflows for model distillation and preference alignment to optimize Small Language Models (SLMs) for efficient, local, resource-constrained inference.
- Languages:
- AI/ML Frameworks:
- Infrastructure:
| Project | Description |
|---|---|
| gpr-slippage-forecasting | A Bayesian framework using Gaussian Process Regression to predict and quantify uncertainty in infrastructure project slippage. |
| project-crimex | A comprehensive analytical platform for crime data visualization and predictive modeling using hybrid ML approaches. |
| varima-nn | Hybrid forecasting system combining Vector ARIMA with Neural Networks for multivariate data. |
| llm-wrappers | A versatile chatbot framework supporting local inference and code interpretation. |
| llm-local-setup | Optimized configurations and scripts for running high-performance local LLM inference and SLM training using CUDA/MPS. |
| cuburt/ai-toolkit | Implementation of DPO and GRPO training loops for aligning Diffusion Transformer (DiT) to specific preferences. |
- Gaussian Process Regression (GPR): Published work on using GPR for project slippage forecasting, providing probabilistic intervals for better project governance.
- Diffusion Scalability: Researching the transition from convolutional denoising to attention-based denoising in latent spaces.
- Tabular & Multimodal RAG: Developing methods to improve LLM reasoning over sparse, large-scale financial tables using semantic chunking.
- LLM Scalability & Efficiency: Investigating quantization-aware training, KV-cache optimization, and distillation techniques to maintain high-reasoning performance in Small Language Models (SLMs) for edge deployment.
- Bio-plausibility in ML: Exploring neural architectures inspired by predictive coding and synaptic plasticity to develop more efficient, lifelong learning systems that mimic biological processing.
"Advancing AI through Bayesian uncertainty, Transformer scalability, and an unhealthy relationship with PyTorch."