Multi-Expert · Memory · No Retraining
Start Exploring Scroll to exploreWe are building a proof-of-concept architecture that demonstrates how AI systems can reason like AGI: multiple specialized experts, persistent memory, world state tracking, and continuous learning—without retraining.
Input flows through perception → memory retrieval → expert modules → planner → action, with a world model and learning loop closing the cycle. Run fully local with Ollama or scale to cloud LLMs.
LLM-based input understanding: intent, entities, domain, complexity.
FAISS vector store for experiences. Semantic retrieval, no retraining.
Structured state: entities, events, goals, relationships.
Specialized reasoning modules. Add/remove at runtime.
LiveBench-optimized coding, VLM for camera and live voice.
Open-Meteo, DuckDuckGo—live data, no API keys required.
Auto-scale experts by query complexity. Unlimited on-demand experts.
Ollama, DeepSeek, OpenAI, Claude, Gemini. Per-expert provider override.
Compared to RAG-only or fine-tuning-only setups, this architecture offers experience-based learning, multi-expert reasoning, and transparent decisions.
Experience memory grows from interactions. No retraining, instant updates.
Math, risk, planning, coding—each expert contributes; planner synthesizes.
See which experts were consulted and how the final answer was formed.
Fully local with Ollama. Optional cloud APIs when you need scale.
This POC demonstrates the core pattern. The scaling path: add more expert modules (100s → 1000s), scale memory to millions of experiences, distribute across GPUs, add domain knowledge bases. The architecture stays the same—you scale the components.
Open-source POC. Private code by Kirill Pokidov. Explore the repo, run locally with Ollama, or reach out for collaboration.