BackRobotics-Based ERP System

Case study · Multi-tenant ERP & POS

Robotics-Based ERP System

Fully automated order ingestion from eBay and Walmart with zero manual entry

Context

A productized, multi-tenant evolution of the dropship pipeline patterns built at Gridiron Tire. Suppliers retain ownership of their products; the ERP handles catalog management, real-time inventory sync, marketplace listing, order processing, and fulfillment automatically. The platform spans a React Native POS, a FastAPI backend, a Next.js integration portal, Django marketplace connectors for eBay and Walmart, an event-driven ERP middleware, MCP servers for AI agents, and a warehouse placement algorithm. All services are containerized and deployed on AWS, with infrastructure defined through Terraform for repeatable environments, plus Sentry and Grafana for observability and GitHub Actions for CI/CD.

Role and scope

Full Stack Software Engineer: designed and built all services end to end, including the POS app, backend API, marketplace integrations, ERP middleware, MCP servers, and infrastructure.

Challenge

Suppliers wanted to sell their own inventory across multiple marketplaces without giving up product ownership or manually managing listings, stock, and orders across each channel. A unified, multi-tenant pipeline was needed to scale this across many suppliers.

Approach: Built a suite of interconnected services: a cross-platform POS (React Native + Expo) with Stripe, barcode scanning, and real-time search via Elasticsearch; a FastAPI backend with multi-tenant OIDC auth, Celery tasks, and SFTP feed delivery; a Next.js API Integration Portal for key management and analytics; Django marketplace connectors for eBay and Walmart with per-tenant OAuth2; an event-driven ERP middleware syncing inventory, e-commerce, and CRM via webhooks; and MCP servers enabling CrewAI agents to orchestrate tasks across all services.

Key decisions and trade-offs

  • Chose multi-tenant services over single-tenant deployments so one platform could serve many suppliers without duplicating infrastructure.
  • Kept marketplace connectors as dedicated Django services so OAuth2 and order ingestion failures stayed isolated from the POS and ERP core.
  • Accepted operational complexity of multiple containerized services in exchange for clearer ownership boundaries and independent deploys.

Outcome

  • Multi-tenant architecture supports many suppliers from a single deployment, each retaining ownership of their inventory
  • Fully automated order ingestion from eBay and Walmart with zero manual entry
  • POS runs on iOS, Android, and web with sub-100ms product search
  • MCP + CrewAI agents handle automated data processing and task execution
  • Repeatable AWS infrastructure provisioned with Terraform and Docker-based service deployments, with Sentry/Grafana observability and GitHub Actions CI/CD

Supporting capability

Additional capability: Voice-assisted ERP operations

As part of the ERP platform, I also built a voice interface that allows operators to ask questions about inventory, warehouse placement, delivery routes, and ERP data. The assistant connects each request to the appropriate business workflow and speaks the response back.

This gives operators a natural way to access operational information without manually searching through multiple systems.

How it works

  1. 1Operator speaks
  2. 2The request is understood
  3. 3The right ERP or operations tool is used
  4. 4The result is prepared
  5. 5The answer is spoken back

What it supports

  • Inventory lookups
  • Warehouse placement information
  • Warehouse utilization checks
  • Delivery route planning
  • ERP data requests
  • Reduced manual searching across operational systems

The broader ERP system connects business software with physical operations, including inventory, storage placement, and delivery planning. The voice interface provides an additional way for operators to access these workflows.

Technology behind this capability

Voice conversation
PersonaPlex
Supports the conversational voice experience.
Speech understanding
faster-whisper
Converts spoken requests into text.
Spoken responses
Qwen TTS
Speaks the response back to the operator.
Wake-word interaction
openWakeWord
Real-time communication
FastAPI and WebSockets
Business workflow coordination
CrewAI
Coordinates broader ERP workflows.
Language-model connectivity
LiteLLM
Connects the workflow to language models.
Connected business tools
Custom MCP servers
Connect the assistant to inventory and route-planning tools.
AI workflow monitoring
LangSmith
Monitors and evaluates request routing, tool usage, failures, timing, and response quality.

Stack

Stack used in this system · Full Stack Software Engineer

FastAPIDjangoPythonReact NativeExpoTypeScriptNext.jsPostgreSQLElasticsearchAWSRedisCeleryDockerTerraformLiteLLMLangSmithWebSocketsCrewAISentryGrafanaGitHub Actions

Architecture stages

  1. 01React Native POS
  2. 02FastAPI backend
  3. 03Django marketplace connectors
  4. 04Event-driven ERP middleware
  5. 05Terraform-managed AWS infrastructure
  6. 06Observability
Mercor Driver Data Platform