Nexo

The Autonomous AI Operations Agent for Modern Commerce

Comment

Inspiration

Modern e-commerce businesses handle a growing volume of customer complaints, but identifying the root cause of an issue can be difficult. Problems may stem from payments, inventory, order management, shipping, or other parts of the customer journey.

Resolving these issues often requires support agents and on-call engineers to manually investigate multiple systems, including payment gateways, inventory databases, order management platforms, and shipping portals. While generative AI has transformed customer-facing chatbots, most LLM solutions stop at providing conversational answers rather than taking action to resolve the underlying problem.

This inspired me to build Nexo around a fundamental question: What if an AI agent could not only identify what went wrong, but autonomously investigate the evidence, execute corrective actions, verify the outcome, and safely escalate when necessary?

What it does

Nexo is an AI operations agent for e-commerce that takes incidents from initial report to verified resolution.

It follows a controlled workflow:

  • Investigate: Collects data from customers, orders, payments, shipments, and inventory.
  • Diagnose & Plan: Uses Google Gemini to identify the likely cause and determine the appropriate action.
  • Validate: Checks proposed actions against prompt-injection protections and business rules.
  • Execute: Performs approved actions such as refunds, cancellations, and inventory adjustments.
  • Verify: Queries the source systems to confirm that the changes were successfully applied.
  • Resolve or Escalate: Emails the customer when the issue is resolved, or creates a Jira incident and alerts engineers through Slack when human intervention is required.

The Nexo Operations Console lets operators monitor the agent's actions, tool calls, policy checks, and audit logs in real time through Server-Sent Events (SSE).

How we built it

Nexo is built with a decoupled architecture that separates AI reasoning from system execution:

  • AI Agent Core: Uses the official Google GenAI SDK (google-genai) with Google Gemini (gemini-3.5-flash-lite) for multi-turn tool use with structured function schemas.
  • Backend & Orchestration: Built with FastAPI (Python 3.12), Redis, and ARQ for asynchronous, long-running investigations.
  • Safety & Verification: Uses a policy engine to validate actions and a post-action verifier to confirm critical changes through read-after-write checks.
  • E-Commerce Layer: Uses a port-and-adapter architecture with an ecommerce_mock microservice that simulates e-commerce failures, edge cases, and webhook events.
  • Operations Console: Built with Next.js 15 and React 19, styled with Tailwind CSS, with live SSE telemetry, state inspection, and audit logs.
  • Deployment: Containerized with Docker and Docker Compose, with deployment support for Google Cloud Run and Cloud SQL for PostgreSQL.

Challenges we ran into

Deployment was a huge challenge. Since this was my first time deploying an architecture this complex, production behaved very differently from running it locally on localhost. With multiple services interacting, I had a really hard time debugging, tracking container errors, and pinpointing where things went wrong.

Accomplishments that we're proud of

I entered the competition past the midpoint with limited time, but I managed to build, connect, and fully deploy a working multi-service autonomous agent from scratch. Even if it's not enterprise-level yet, making the entire closed-loop system, from Gemini reasoning to real-time verification and dashboard streaming, actually work end-to-end is something I'm genuinely proud of.

What we learned

I learned that building an autonomous agent is only half the battle, the harder part is engineering deterministic guardrails so the model doesn't hallucinate actions, and orchestrating the backend infrastructure so multiple services can actually communicate reliably in production.

What's next for Nexo

  • Real-World Integrations: Swap out the mock e-commerce platform with production API adapters (Shopify, Stripe, WooCommerce, EasyPost).
  • Industry-Grade Reliability: Harden the architecture with automated retry policies, distributed tracing, and production-grade monitoring to make it true enterprise-ready.
  • Human-in-the-Loop Safeguards: Add one-click operator approvals for high-value financial actions before the agent executes them.

Built With

Share this project:

Updates

Submission history