OpsPilot AI

Inspiration Small business owners and e-commerce operators lose hundreds of hours every month dealing with repetitive, judgment-heavy operational friction: inspecting damaged-goods claims against return policies, reconciling supplier invoices against Purchase Orders, and managing customer inquiries.

Standard conversational AI chatbots are ineffective because they merely chat without executing real operations. On the other hand, unconstrained autonomous agents pose serious financial liability by issuing unauthorized refunds or modifying records without oversight. We built OpsPilot AI to deliver real operational work with strict, deterministic human-in-the-loop safety boundaries.

What It Does OpsPilot AI is an autonomous operations teammate that handles operational workflows end-to-end:

Damaged-Order Complaints & Warranties: Ingests customer complaints, uses Amazon Bedrock Multimodal Vision to inspect damage photos (96.8% transit damage verification), retrieves customer and order histories from DynamoDB, checks SOP warranty rules, and stages remediation. Strict Human-in-the-Loop Safety Gate: Routine actions ($\le $50$) execute autonomously. High-risk operations ($> $50$ refunds, legal dispute replies) halt at the APPROVE stage and queue in the Approval Center for supervisor sign-off. Vendor Invoice Reconciliation: Cross-references invoices against Purchase Orders, validates variance ($< 1%$), and schedules payments. Cryptographic SHA-256 Audit Trail: Commits every tool call and supervisor decision into an immutable, sequenced hash chain in Amazon DynamoDB. Cross-Platform Responsive UX: Works on mobile, tablet, and desktop with an off-canvas drawer and bottom navigation bar. How We Built It Orchestration: Built using the Python Strands Agents SDK with custom @tool decorators (get_customer, get_order, search_business_policy, analyze_document, send_email, issue_refund, create_followup_task, record_audit_event). Foundation Models: Amazon Bedrock running Anthropic Claude 3.5 Sonnet (anthropic.claude-3-5-sonnet-20241022-v2:0) and Amazon Nova for reasoning and multimodal damage inspection. Cloud Infrastructure: Serverless deployment via AWS CDK (AWS Lambda, AWS API Gateway v2 HTTP API, Amazon DynamoDB, Amazon S3 with KMS encryption, and Amazon SES). Frontend: React 19, TypeScript, Vite, Tailwind CSS, with dual-mode capability (zero-AWS demo simulation vs. live AWS cloud connection). Challenges We Ran Into Prompt Injection Defense: Customer emails and invoice OCR text can contain malicious instructions attempting to bypass return limits. We solved this by wrapping all external data in an isolated envelope and instructing the agent to treat it strictly as inert data to extract facts from. Deterministic Approvals: Ensuring the agent cannot bypass approval gates even if it believes the customer deserves the refund. Accomplishments We're Proud Of A live, working 9-stage transparent pipeline visualizer (UNDERSTAND $\to$ PLAN $\to$ RETRIEVE $\to$ REASON $\to$ DECIDE $\to$ APPROVE $\to$ EXECUTE $\to$ VERIFY $\to$ AUDIT). Instant Demo Mode enabling hackathon judges to test the entire solution immediately without requiring AWS credentials or cloud setup. 100% compliant with Hackathon rules, including an open-source MIT License. What We Learned We learned how the Strands Agents SDK cleanly separates tool definitions from model orchestration, making Amazon Bedrock foundation models fast and safe for enterprise automation.

What's Next for OpsPilot AI Adding multi-agent Swarm collaboration for logistics dispatch and direct ERP connectors for Shopify, QuickBooks, and SAP.