Inspiration
A shift handover should distinguish a recorded action from a resolved problem. SevaRelay explores that distinction for a rural service-desk workflow, using fictional cases rather than operational customer records.
What it does
The local web app turns a six-case queue into an inspectable next-action workflow. An operator can read the source notes, review the missing evidence, preview an English or Hindi desk note, cancel it or explicitly confirm it, and generate a handover brief for the next operator. Recording a reminder does not silently mark a missing document or receipt copy as received. Nothing is sent to a customer.
How it was built
This is an AI-assisted build and a clearly labeled deterministic Alexa+ simulation. OpenAI assistants helped create and independently review the prototype's code and documentation. The running application uses a small intent router, templates and explicit buttons, not a runtime language model.
A dependency-free Node.js service implements a limited MCP 2025-11-25 Streamable HTTP server. The custom browser client performs initialization, tool discovery and actual tool calls; a visible protocol trace connects the interface to the requests and responses. The entire demonstration runs on loopback with synthetic fixtures and no outbound service calls.
Challenges and verification
The main challenge was preserving correct state across previews, cancellations, repeated confirmations and interrupted requests. Versioned previews, expiration checks and repeat-safe confirmation protect the local workflow from stale context and repeated clicks. Independent QA found and fixed eight defects. The final build passes 57 automated tests, a real MCP smoke workflow and a fresh browser check covering preview, cancel, confirm, Hindi output, filters, handover, reset and interrupted-server recovery. These checks do not establish production reliability or complete MCP conformance.
What is distinctive
The interface makes evidence and human approval visible. It separates unresolved evidence, recorded desk notes and ready cases, so a logged action cannot be mistaken for an actual service outcome.
Limitations and next steps
SevaRelay is a reproducible prototype, not a deployed business service. It has no production authentication or persistent database, no real users or measured impact, and no customer-contact, payment or live Alexa integration. A real deployment would require operator research, service-specific requirements, multilingual usability testing, secure identity and storage, and separately authorized integrations.
Tool disclosure
Built with OpenAI assistant coding and research support, Node.js, HTML, CSS and JavaScript, plus the project's own MCP implementation. No Kiro, AWS, Bedrock, paid API, Amazon device or live Alexa+ tooling was used. Atul Bhat is the individual entrant; AI assistants are not human teammates.
Log in or sign up for Devpost to join the conversation.