The Engine Room
Three invariant subsystems. Purpose-built for deterministic, self-healing AI execution.
Hybrid Inference Engine
Dynamically routes inference requests between local GPU compute and cloud endpoints. Detects available VRAM at boot, then commits to the optimal path.
- Local: LiteLLM → GPU (≥8GB VRAM)
- Cloud: Managed API fallback
- Routing: Deterministic, zero-latency switch
Vector Memory
Persistent, semantic retrieval layer backed by Pinecone. Every Skill's knowledge base is embedded at 1536 dimensions and queried at inference time for RAG context.
- Engine: Pinecone (1536-dim cosine)
- Embedding: text-embedding-3-small
- Namespacing: per-skill isolation
Atomic Stability
Every Skill executes inside a containerised sandbox. BTRFS snapshots provide instant rollback. A failed Skill never corrupts the host environment.
- Runtime: Docker container isolation
- Snapshots: BTRFS atomic rollback
- Kriya Loop: self-annealing on error
Skill Store
Deployable agent capabilities. Each Skill runs as an atomic unit inside the Kriya Loop.
Python Environment Architect
Automatically provisions, audits, and repairs Python virtual environments. Resolves dependency conflicts using RAG-retrieved compatibility matrices.
Video Render Auto-Pilot
Orchestrates end-to-end video render pipelines. Manages frame queuing, codec selection, and hardware GPU scheduling for local-first processing.
System Log Optimizer
Ingests and semantically indexes system logs. Detects anomaly patterns via vector similarity, then proposes self-healing patches through the Kriya Loop.
Deep Research Agent
Multi-step, web-grounded research synthesiser. Fetches sources, embeds into Pinecone workspace, then delivers a structured intelligence brief.