Leonardo Cofone

DSALT: Dynamic Sparse Attention with Landmark Tokens DSALT is a research project on Transformer attention mechanisms that investigates the structural limitations of dense self-attention in long-context settings. In particular, it analyzes how uniform softmax weighting can lead to cumulative representational noise and progressive homogenization of token embeddings across layers. To address this, DSALT introduces a dynamic sparse attention mechanism that combines local windowed attention with globally selected landmark tokens. Landmark selection is performed using a hybrid energy-based scoring function that balances representational magnitude and output relevance. The objective is to reduce redundant token interactions while preserving long-range dependencies, improving both representational quality and computational efficiency. The project also explores the relationship between attention structure, rank dynamics, and information flow in deep neural networks.

LerriAI is built around a simple idea: most people don’t lack ability, they struggle with organization. Life throws countless tasks, deadlines, emails, routines, goals, and unexpected problems at us, and even the most motivated individuals can easily feel overwhelmed. LerriAI exists to manage that complexity. It is not just a chatbot or a simple assistant, it is a living system of intelligent agents that work together to structure your day, support your decisions, and keep your personal world running smoothly. Each agent has a specific role, analyzing your habits, priorities, and preferences, planning, reminding, optimizing, summarizing, and tracking progress, while communicating through a shared intelligence layer. In this way, LerriAI turns chaos into order, complexity into clarity, and everyday distractions into focused, actionable guidance, helping you stay on track with what truly matters.