INT8 tensor subclass for PyTorch enabling up to 4× memory reduction with optimized CUDA/Triton kernels. Merged into TorchAO.

Benchmarking vLLM, SGLang, and HuggingFace TGI across agentic workflows, long-context, and high-throughput scenarios.

Minimalist implementation of vLLM for educational purposes and rapid prototyping of inference optimizations.

Improving Visual Transformer robustness using occlusion generators for better generalization.

Building sparse frame selection methods for faster video reasoning and efficient temporal understanding.

Using LLMs for prediction market analysis and probabilistic forecasting.

Developing LLM-powered trading agents that react to market events and news.