I built an AI orchestration system where each component maps to a brain region. Not as a metaphor — each region is a separate module that influences how queries are actually processed.
The architecture has 12 regions:
- Hippocampal Memory — tracks which strategies worked for which query types
- Prefrontal Monitor — assesses confidence before committing to a response
- Amygdala — detects emotional tone (frustrated, curious, confused) and calibrates response style
- Dorsal Attention Network — allocates token budget and model count based on query complexity
- Cerebellum — plans parallel vs sequential model execution, implements circuit breakers
- Insula — monitors error rate, latency, token burn, and disk pressure; auto-throttles when stressed
- Corpus Callosum — HTTP API bridge between two hemispheres (user-facing service + autonomous worker)
- Temporal Pole — builds user profiles over time (expertise, domain interests, communication style)
- Plus 4 more (Association Cortex, Synaptic Plasticity, Basal Selection, Cingulate Detector)
Every query passes through a 21-step pipeline. Multiple models (GPT, Claude, Gemini, Mistral, Grok) are queried in parallel, validated for novelty and coherence, then synthesized.
Key design decisions:
1. All brain modules are rule-based heuristics, not LLM calls. Zero additional API cost. The Amygdala is pattern matching on keywords and sentence structure, not calling GPT to detect emotion.
2. Every module is wrapped in try/except. If any brain region fails, the query still completes normally. They add intelligence but never block.
3. Two hemispheres run as separate services. The "Left Brain" handles user requests. The "Right Brain" runs 24/7 autonomously researching global problems. They share discoveries through the Corpus Callosum (an internal API bridge).
4. The Insula implements genuine interoception — it monitors the system's own health metrics and throttles when stressed (reducing model count and iterations), similar to how biological interoception regulates behavior under physiological stress.
What it's NOT: This isn't AGI. The brain mapping is an architectural pattern for organizing multi-model orchestration, not a claim about consciousness or general intelligence. The neuroscience analogy helps structure a complex system in a way that's intuitive and extensible.
Stack: Python/FastAPI backend, Next.js frontend, SQLite for each brain region's state, deployed on Render.
aegismind_app•1h ago
The architecture has 12 regions:
- Hippocampal Memory — tracks which strategies worked for which query types - Prefrontal Monitor — assesses confidence before committing to a response - Amygdala — detects emotional tone (frustrated, curious, confused) and calibrates response style - Dorsal Attention Network — allocates token budget and model count based on query complexity - Cerebellum — plans parallel vs sequential model execution, implements circuit breakers - Insula — monitors error rate, latency, token burn, and disk pressure; auto-throttles when stressed - Corpus Callosum — HTTP API bridge between two hemispheres (user-facing service + autonomous worker) - Temporal Pole — builds user profiles over time (expertise, domain interests, communication style) - Plus 4 more (Association Cortex, Synaptic Plasticity, Basal Selection, Cingulate Detector)
Every query passes through a 21-step pipeline. Multiple models (GPT, Claude, Gemini, Mistral, Grok) are queried in parallel, validated for novelty and coherence, then synthesized.
Key design decisions:
1. All brain modules are rule-based heuristics, not LLM calls. Zero additional API cost. The Amygdala is pattern matching on keywords and sentence structure, not calling GPT to detect emotion.
2. Every module is wrapped in try/except. If any brain region fails, the query still completes normally. They add intelligence but never block.
3. Two hemispheres run as separate services. The "Left Brain" handles user requests. The "Right Brain" runs 24/7 autonomously researching global problems. They share discoveries through the Corpus Callosum (an internal API bridge).
4. The Insula implements genuine interoception — it monitors the system's own health metrics and throttles when stressed (reducing model count and iterations), similar to how biological interoception regulates behavior under physiological stress.
What it's NOT: This isn't AGI. The brain mapping is an architectural pattern for organizing multi-model orchestration, not a claim about consciousness or general intelligence. The neuroscience analogy helps structure a complex system in a way that's intuitive and extensible.
Stack: Python/FastAPI backend, Next.js frontend, SQLite for each brain region's state, deployed on Render.
Try it: https://www.aegismind.app
Happy to discuss the architecture decisions or answer questions about the design.