Capabilities
What Doradus operates today. Each capability listed here is a live system serving production traffic — not a roadmap item.
For the full product surface (interactive consoles, build pipelines, per-tenant integrations), see perfct.ai.
Model fleet
Tiered model serving across vLLM, SGLang, and Mac MLX hosts. Category-routed (AGENTIC_FRONTIER, VISUAL_FRONTIER, FORMAL_PROOF, FRONTIER_HIGH_CONTEXT, etc.) with prefix-cache anchors and fail-closed scarce-tier reservations. Cross-host coordination via Consul DNS round-robin and per-host LRU; no central optimizer.
AgentFlow pipelines
Lite/heavy execution lanes routed through a single pipeline_routing config. Every LLM call goes through the centralized routing catalog — no direct vendor SDK invocations from skills or modes.
BMAD modes
Multi-mode pipeline (dev, brownfield, design, vision, apple-native) with diffusion visual-verification, refrag image namespace ingestion, and per-mode model-category reservations.
Swap-pool architecture
Per-host llama-swap pools across the 4×32GB and 5×96GB GPU tiers. Auto-load on first request, idle TTL eviction, and a predictive warm-hint autoscaler. OpenAI-compatible API surface.