Inspiration
Small business owners should be able to ask what a visibility report means and what to do next without learning an analytics dashboard. We built a conversational demonstration of that journey for the Alexa+ simulation route. It supports Balasar's work on websites and AI visibility for Dutch and Turkish businesses; the present demonstration is in English and uses fictional businesses.
What it does
An owner chooses a fictional business, asks how visible its website is, and receives an explanation grounded in a fixed demonstration snapshot. A second question turns the snapshot into practical priorities. A request for human help leads to an explicit confirmation step and creates a local follow-up draft. It does not send a message or create a real customer record.
The demonstration also exposes its limits. It does not perform a live website scan, promise search rankings, or reveal another business's information. A deliberately hostile page title is treated as page data, rather than permission to change the assistant's behaviour.
How we built it
The application uses React and TypeScript with deterministic business fixtures and tool routing. Its optional loopback server keeps session authorization and provider access outside the browser. The selected business remains fixed for that session. Local tools provide facts and permitted actions; a model does not authorize follow-up or invent measurements.
An optional Amazon Bedrock Converse adapter targets Amazon Nova Micro. It includes an explicit spending window, a call limit, no automatic SDK retries, and an audit record without raw prompts or answers. A claim check can replace unsupported model wording with a deterministic response. The separate Bedrock trial is BLOCKED_COST_GATE: the budget check was denied (budgets:ViewBudget), and zero model calls were made. This submission uses deterministic mock mode and does not enter the AWS Builder mini challenge.
Challenges and lessons
Separating factual content, wording and permission to act made the workflow easier to inspect. Windows file handling also required separate verification; a successful Linux run did not prove Windows behaviour. We retain explicit gaps instead of treating one test as a general security guarantee.
What is demonstrated
A multi-step visibility conversation, priorities, consent before creating a local follow-up draft, refusal paths, and handling of hostile fixture text. Tests cover the conversation engine and bounded provider path. This is an unofficial Alexa+ web simulation using fictional data. It is not connected to a real Alexa+ device or runtime, and the in-process tool interface is not a fully compliant deployed MCP server. The English demonstration video is approximately 2 minutes 22 seconds long and uses an AI-generated ElevenLabs Noah voice-over. The final standalone export has completed an independent review for local demonstration. Two HTTP tests remain recorded as failed in the review environment; a separate loopback check verified rejection of invalid Host headers. The owner has approved the narration and pronunciation.
What's next
Any future live Bedrock test requires separate budget authorization and verification; it is not needed to run this mock demonstration. Real visibility data, Dutch and Turkish language support, customer deployment and real follow-up integrations require separate work and approval.
Log in or sign up for Devpost to join the conversation.