Skip to main content

ObVect

The AI backend and graph compiler everything else in Colossal Capital runs on.

Type: AI Backend + Graph Compiler · Category: AI & Dev Tools

Open ObVect

Managed AI backend and product graph compiler — one API for inference, retrieval, agents, training, and gap detection behind Akashic, TopOS, and products across the ecosystem. OmniRoute, the first-class model gateway, scores each request on cost, availability, model fit, context window, and time-to-ready, then places it across Together AI, HuggingFace TGI, OpenRouter, RunPod, Anthropic, OpenAI, DeepSeek, and more — with model tiers, free-mode pools, BYOK/BYOM, failover, and audit logs. The Virtuo SDK turns any model into a multi-turn agent — models without a native loop get one — and every agent reads and writes one shared reasoning graph, with human-in-the-loop approval gates tied to RBAC inside the graph itself. The Ladder graph compiler builds a 7-layer product graph from connected repos, detects eight classes of structural gap, evaluates Product Board health, and generates ordered tickets in contract-first dependency order — Terraform for products. Neo4j knowledge graphs and Milvus hybrid search ground answers in your code and docs. REST and MCP expose the same surface for portals, IDEs, and automations.

Highlights​

  • Smart inference routing — Every request scored on cost, availability, model fit, context window, and latency through the OmniRoute gateway — logged so you can audit why a provider won.
  • Virtuo reasoning graph — Every agent is wrapped in a multi-turn loop if it is not already loop-capable, and they all share one reasoning graph — extraction, strategy, tool execution, observation, decision, reflection nodes. HITL approval gates are built into the graph and bound to RBAC.
  • Confidence-gated HITL — Every agent decision carries a confidence score from the reasoning graph. High-confidence steps auto-approve; low-confidence steps pause for human review. Agents ask for help exactly when they should.
  • Global kill switch — One button stops every agent across the platform — every loop, dispatcher, sandbox, and assistant. Enforced at both the execution layer and the Idio PDP for defense-in-depth.
  • Unified explain endpoint — Trace any agent interaction back to its reasoning map, audit trail, provenance marks, and cost in one API call. Every decision, handoff, and token joined into a single chronological timeline.
  • Real-time agent control — Pause, resume, or cancel any running agent or agent loop from the portal dashboard. Cooperative control checks flags between turns — no orphaned processes, no runaway loops.
  • Ladder graph compiler — Builds a 7-layer product graph from connected repos, detects eight classes of structural gap (missing outcomes, orphaned nodes, cross-repo duplicates, redundant paths, stack layer holes, under-utilized personas, gate coverage holes, missing human interactions), evaluates Product Board health with cross-repo duplication and consistency analysis, and generates ordered tickets in contract-first dependency order — Terraform for products.
  • Grounded retrieval — Neo4j knowledge graph plus Milvus BM25 + vector hybrid search for tenant-scoped RAG.
  • MCP + REST — Same auth and routing for Cursor, VS Code, product portals, and external agents.
  • Training flywheel — Run telemetry feeds SFT/DPO LoRA fine-tunes; contextual bandit and W&B loops promote improvements with HITL review. Per-tenant private training with NUF (Unclonable Friend) models that eventually replace third-party API keys.

Features​

  • Inference — Chat, completion, and structured analysis — sensible defaults with per-request model overrides.
  • Retrieval — Neo4j KG plus Milvus BM25 + vector hybrid search for grounded RAG across tenant-scoped corpora.
  • Ladder graph compiler — Multi-repo gap detection, Product Board health evaluation, constraint-based design, and ordered ticket generation — the graph → analyze → ticket → agent → code → re-analyze loop.
  • MCP server — SSE endpoint for Cursor, VS Code, and external agents — same auth and routing as REST.
  • Agent runtime — Virtuo SDK multi-turn loops, shared reasoning graph, and RBAC-bound human-in-the-loop gates with confidence-based auto-approval.
  • Agent safety — Global kill switch, confidence-gated HITL, unified explain endpoint, and real-time pause/resume/cancel control.
  • Training — Telemetry-driven LoRA fine-tunes with HITL promotion — models improve from real usage. NUF private models, cost summary, and model comparison.

Tags​

Virtuo SDK · MCP · OmniRoute · Inference Router · Reasoning Graph · Ladder · Graph Compiler · Milvus · Neo4j · Training · Confidence-Gated HITL · Kill Switch · Explain Endpoint · Agent Control