THE BLACK EYE MEDIAStart a Project
← Back to portfolio

AI Operations

Advanced AI Operations Platform

UAT closeout PASS · M0–M10

LangGraph agents, MCP tool gateway, Qdrant RAG, HITL, and JWT admin RBAC — n8n for side effects.

M0–M10 complete · API 0.10.0 · Closeout 2026-09-11

The problem

Ops automation is fragmented, tools are uncontrolled, and high-risk writes go live unchecked.

  • Fragmented ops automation

    Email, finance, CRM, and ERP each have one-off scripts with no shared brain.

  • Ungrounded LLM answers

    Policy/pricing invented; ops teams lose trust.

  • Uncontrolled tool use

    Agents call APIs without role allowlists or audit.

  • High-risk writes unchecked

    Invoices, quotations, outreach without HITL packages.

  • Silent failures

    Workflows die; duplicates reprocess; no DLQ or metrics.

The solution

FastAPI + LangGraph own AI/MCP; n8n owns thin side effects — audited tool use and HITL on high-risk writes.

A Dockerized AI workforce layer for sales, finance, and ops: shared orchestrator, tool permissions, grounded answers, and an operator cockpit — not a chatbot bolted to scripts.

  • LangGraph orchestrator

    Multi-agent routing with shared memory and guardrails.

  • MCP tool gateway

    Allowlisted tools with audit — CRM, ERP, research, docs.

  • Qdrant RAG

    Grounded policy/product answers with confidence gates.

  • HITL packages

    High-risk writes require structured approval before live.

  • Admin SPA RBAC

    Approvals, runs, eval, and health in one JWT cockpit.

Triggers → FastAPI/LangGraph → MCP tools + RAG → Postgres/Redis/Qdrant → n8n side effects → HubSpot/Odoo/notify

Before → After

AI Ops Platform before and after

Architecture

AI Ops Platform architecture diagram

Delivery

Milestone narrative from kickoff through closeout — all gates met.

IDThemeUnlocked
M0–M1Architecture + skeletonCompose stack, schema, FastAPI, n8n, health
M2–M3Core intelligenceOrchestrator/agents, RAG, primary integrations
M4–M5Reliability + HITLApprovals, DLQ, audit, Phase 1 gate
M6–M7Depth + Phase 2Secondary channels/agents, Phase 2 regression
M8–M9Dashboard + scaleOps UI RBAC, eval, observability, adapters
M10CloseoutDocs, UAT/smoke, KT pack, deliverables

Proof

Verified on local Docker dry-run — no invented client metrics.

  • Closeout

    PASS

    Phase 1–3 gates across uat_closeout

  • Milestones

    M0–M10

    Complete with documented adapters

  • HITL

    Gated writes

    High-risk finance/outreach require approval

  • API

    0.10.0

    Ops platform release

Stack

  • FastAPI
  • LangGraph
  • MCP
  • n8n
  • Postgres
  • Redis
  • Qdrant
  • OpenAI
  • HubSpot / Odoo
  • Caddy
  • Prometheus / Grafana

Want a system like this?

We build production AI ops platforms — multi-agent orchestration, RAG, CRM, and reliable automation — not one-off chatbot demos.