Yeti Nepal v1: 500M Parameter Hybrid Model
A hybrid Mamba-2 SSM, Grouped-Query Attention (GQA), and Mixture-of-Experts (MoE) causal language model optimized for free-tier Google Colab GPU training.
Offline AI Brain Architecture
An in-depth exploration of designing AI systems that operate entirely without internet connectivity — covering memory layers, local reasoning engines, retrieval pipelines, and the trade-offs of local-first AI design.
Multi-Agent Coordination Patterns
Researching how multiple AI agents can coordinate tasks, share memory, and avoid conflicts — with real architecture patterns and failure modes documented.
Local Knowledge Retrieval Systems
How to build fast, private, and accurate knowledge retrieval systems that run entirely on local hardware without vector cloud services.
See the projects behind the research
Most research articles are backed by working implementations. Explore the projects section to see the code.