BOLT for Alexa+

Inspiration

AI assistants are becoming useful for everyday operational tasks, but a plausible recommendation is not the same as permission to take action. A shipment may look delivered in one system and delayed in another. BOLT for Alexa+ explores a simple human-first question: can a conversational assistant show the evidence, prepare a draft, and pause before doing anything consequential?

What we built

A working, browser-based simulation of an Alexa+ conversational experience: a safe review interface for a fictional shipment, order U-1042. It is entirely local, with synthetic data and no account required.

A user can:

• Ask what happened to the order and see conflicting observations from OMS, warehouse and carrier snapshots.
• Request a summary draft, clearly marked as unsent.
• Request a customer notification or priority change, which triggers an explicit human approval gate tied to a unique request ID.
• Approve or reject that request. Approving generates a simulated receipt only; no message or status change takes place.
• Try a destructive command, which is refused under a default-deny rule.
• Review and export the in-memory decision log as JSON.

Try the interactive working demo: https://ondmindmanagement-hub.github.io/bolt-alexa-hackathon/

Watch the 65-second captioned browser demonstration: https://ondmindmanagement-hub.github.io/bolt-alexa-hackathon/demo/Unfire_BOLT_Alexa_Simulation_2026.mp4

Source and reproducibility: https://github.com/ondmindmanagement-hub/bolt-alexa-hackathon

How we built it

The application uses vanilla JavaScript, HTML, CSS, and Node.js. A deterministic classifier converts written example utterances into bounded intents: read-only, draft-only, approval-required, or deny. A small state machine binds a local review to an exact pending request reference; unmatched confirmations, duplicate decisions and prohibited operations are blocked. The interface displays the three synthetic evidence records and review history.

Run npm test for 15 simulator tests and three original governance tests. We also tested the five-step workflow in an actual Chrome browser. In that test, no calls to external hosts were recorded.

The app is free to test with no credentials, Amazon account or API keys.

Use of Alexa+ and product feedback

We chose the hackathon's simulated Alexa+ experience route. Users type sample voice transcripts into an Alexa+-inspired web interface; there is no microphone or voice recognition. We did not use an Amazon Alexa+ SDK, MCP server, Agent Skill, production Alexa+ runtime, or an actual LLM in this version. The simulation is intended to test the human-approval interaction before attempting a real platform integration.

What worked well: the simulated route allowed us to build and test a full user journey without device access or credentials; local interaction is fast and reproducible.

What needs further work: real Alexa+ onboarding, authenticated voice identity, production intent handling, tool invocation, durable audit trails and robust reviewer authentication. We would use actual Amazon tooling for a later prototype once access and a suitable non-sensitive use case were established. We have not evaluated its SDK onboarding or runtime reliability and cannot claim direct feedback about technologies we did not use.

Challenges and lessons

The key challenge was preserving a smooth dialogue without silently transforming a recommendation into an action. We learned to treat uncertainty explicitly: conflicting source evidence is shown, a draft stays a draft, and a human can refuse an external action. Simulated approval never claims that a real downstream effect occurred.

What changed during this hackathon

The original repository had a basic three-rule command-line scaffold. It was significantly expanded during the submission window with the complete responsive browser interface, contextual synthetic evidence, five test scenarios, exact-request approval and rejection, a local review log and JSON export, 15 additional automated tests, a reproducible demonstration video, and public GitHub Pages deployment.

Current limits

This is an experimental simulated Alexa+ UX with no external system actions. It does not connect to Amazon, send emails, change logistics records, make payments, authenticate users, or run a live AI model. Synthetic outcomes are not evidence of commercial savings or production reliability. The audit log is ephemeral, not immutable.

Project by Omar Baró · Unfire — https://unfire.technology

Current demo video: The embedded YouTube video is the publicly published, 65-second captioned BOLT for Alexa+ browser walkthrough from the official Unfire channel: https://youtu.be/3PxW3vQEyc0. It demonstrates simulated workflows only, without an actual Amazon/Alexa+ connection.

Built With

• alexa+
• css3
• html5
• human-in-the-loop
• javascript
• node.js
• simulation