Skip to content

Enterprise AI · Mobility & Ride Hailing (UAE)

Tanzim: AI Driver Support & CRM Automation

AI support platform resolving driver tickets across 5-6 workflows, with autonomous email replies with confidence-based escalation into HubSpot CRM.

Discuss a similar project
Role
AI Engineer
Built the agent architecture, guardrail layer, CRM integration, MCP server and the full auth model end to end.
Industry
Mobility & Ride Hailing (UAE)
Timeline
Jun 2026 - Jul 2026

Business impact

agent workflows
3 + email routing
graph nodes
24
mcp tools
16 allow listed
guardrail boundaries
3
auth lanes
4

Overview

AI support platform where drivers raise tickets across 5 to 6 workflows: lost-and-found items recovered in vehicles, fare-dispute resolution and more. Each inbound email triggers an AI workflow that attempts autonomous resolution by drafting and sending replies.

When a case exceeds its confidence or scope, the agent escalates and auto-creates a HubSpot CRM ticket for a human agent to close the loop. Workflows are engineered in FastAPI and orchestrated with LangChain and LangGraph, with layered guardrails over every automated action, hand-off and escalation. Hardened for production: an Azure API Management gateway fronts the AI app, Container Apps run inside a private VNet, and the MCP server is deployed under the hood of the same VNet and exposed only through the API gateway.

Technical architecture

Frontend
React and TypeScript SPA over HTTP/SSE - analyst chat, support chat, CRM grid, trace monitor
Gateway
APIM policy with streaming safe forwarding, CORS allow list, anti spoof header stripping and per IP throttling
Agent Layer
LangGraph workflows with a strict one way hierarchy: workflow, nodes, agents, tools
Governance
Input sanitizer, output reply guard, SQL policy node, tool registry and MCP operation allow list
CRM
HubSpot service owning contacts, ticket CRUD and pipelines, with typed errors and graceful degradation
Identity
Entra ID session JWT, optional TOTP, service key lane and APIM delegated identity bridge
Testing
Hermetic harness faking only external boundaries - guardrail, graph, e2e, multi turn and concurrency suites

Key features

  • Natural language data analyst - router, schema fetch, table selection, SQL generation, policy gate and execution
  • Lost & Found agent that resolves the trip, records the item and links a match when it finds one
  • Fare dispute sub graph with surge, policy and slot checks, supporting multi turn escalation
  • Automatic CRM ticket creation when a lost item report finds no match, associated to a resolved contact
  • Full HubSpot ticket lifecycle - pipelines, stages, priority, update and archive - with the token held server side
  • Prompt injection sanitization on every inbound message before it reaches an LLM node
  • Outbound reply guard that strips refund promises, formatting and emoji, and guarantees the ticket reference
  • Guardrail audit trail - every protection that fires is recorded as a flag on graph state
  • Read only SQL policy gate rejecting eight destructive keywords before execution
  • Native MCP server exposing an allow listed tool set to external AI clients under per user API keys
  • Four authentication lanes plus opt in TOTP two factor with encrypted secrets and recovery codes
  • Live trace monitor surfacing every graph step and tool call to the operator

Challenges solved

  • Preventing the agent from promising refunds - sentence level scrubbing tuned to preserve legitimate wording
  • Prompt injection on a customer facing channel - deterministic sanitization that excises the injection clause only
  • Letting an agent write to a production CRM - scoped trigger, exception safe, and always attributable to a contact
  • Exposing agent tools to third party AI clients where the identity provider has no dynamic client registration
  • Enforcing two factor on a non interactive lane by proving it through the token claim instead of a prompt
  • Text to SQL against a live database - staged generation for accuracy, read only policy gate for safety

Tech stack

  • Python
  • FastAPI
  • LangGraph
  • LangChain
  • Azure OpenAI
  • Model Context Protocol (MCP)
  • HubSpot CRM API
  • React
  • TypeScript
  • Microsoft Entra ID
  • TOTP 2FA
  • Azure API Management
  • Azure Container Apps
  • Text to SQL
  • pytest
  • Docker
All projects