How ahb-sjsu/agi-hpc works
Atlas AI is a distributed cognitive architecture: user requests enter its server, are routed through an ensemble of cloud and local reasoning voices over NATS, and produce responses informed by memory and safety checks. Separate branches support Council deliberation, Erebus learning, Primer tutoring, and dream consolidation. The README documents these capabilities and the shared bus; the source excerpts are mostly from vendored libraries, so implementation-level links among Atlas subsystems are not verified here.
Main entry points: atlas-rag-server.py Atlas request server · agi vMOE ensemble · memory Memory services · service.py Primer tutor
Components 17
Request and routing
- Atlas request server
atlas-rag-server.py - Adaptive routing
src/agi/metacognition/adaptive_router.py
Cognitive capabilities
- vMOE ensemble
src/agi - Frontier models
- Local specialists
- Divine Council
src/agi/reasoning - ErisML safety
src/agi/safety/erisml/service.py - Erebus learning
src/agi/autonomous - Primer tutor
src/agi/primer/service.py - Dream consolidation
src/agi/dreaming/nats_service.py - GPU burst workers
Shared workspace
- NATS JetStream
Memory and learning
- Memory services
src/agi/memory - PostgreSQL + pgvector
- Knowledge graph
src/agi/memory/nats_service.py - Sensei wiki
Other components
- User
Connections 23
- User → Atlas request server: submits request
- Atlas request server → Adaptive routing: routes request
atlas-rag-server.py - Adaptive routing → vMOE ensemble: invokes vMOE
- vMOE ensemble → Frontier models: fans out
- vMOE ensemble → Local specialists: fans out
- vMOE ensemble → NATS JetStream: publishes turns
- NATS JetStream → vMOE ensemble: dispatches messages
- Frontier models → NATS JetStream: joins via leaf
- Local specialists → NATS JetStream: subscribes
- Atlas request server → Divine Council: handles decisions
atlas-rag-server.py - Divine Council → NATS JetStream: deliberates on
- Atlas request server → ErisML safety: handles policy gates
- Atlas request server → Erebus learning: handles puzzles
- Erebus learning → Primer tutor: requests tutoring
- Primer tutor → NATS JetStream: teaching cycle
- Primer tutor → Sensei wiki: publishes verified notes
- Atlas request server → Dream consolidation: supports consolidation
- Dream consolidation → NATS JetStream: uses shared fabric
- Memory services → NATS JetStream: subscribes and publishes
- Memory services → PostgreSQL + pgvector: reads and writes
- Memory services → Knowledge graph: tracks taught and unanswered
src/agi/memory/nats_service.py - Atlas request server → User: returns result
- NATS JetStream → GPU burst workers: dispatches workload