Prepared September 8, 2026 from product-feedback.md for live Devpost fields
28303–28307. The event entry was submitted September 8; these are editable local
reference copies, not a live view of the submitted answers. Subsequent repository
edits do not automatically update Devpost. Observations refer to dated project
tests, not a new AWS audit or unperformed official Alexa testing.
Alexa+ add-on documentation informed the repair-information prototype, the MCP integration route, voice/visual design and confirmation boundaries. The MCP TypeScript server and client SDKs implement actual Streamable HTTP tool calls, typed schemas and transport tests. A custom browser simulator demonstrates the experience with optional browser speech input and typed-command fallback; it is not Amazon's official simulator or a certified add-on.
Amazon Bedrock Converse with Nova Lite provides an optional, non-authoritative comparison lead sentence through a separate IAM-authenticated Lambda/API Gateway narrator. DynamoDB keeps a finite model-attempt allowance. The customer staging site uses Login with Amazon for website identity, API Gateway/Lambda for HTTPS hosting, DynamoDB for sessions, trusted links, enrollment and fictional repair projections, and Secrets Manager for private runtime configuration. CloudFormation, IAM, KMS and finite-retention CloudWatch logs support deployment and authority boundaries. Hosted customer fixtures are separate from local shop state.
TypeScript/Node.js, Zod and Express implement application and validation code; pnpm manages the monorepo; Docker and GitHub Actions run repeatable checks. No AgentCore, Kiro Crew, Strands or official Alexa device/toolkit execution is claimed.
Alexa+ documentation helped separate conversational capabilities, category contracts and consumer-facing policy requirements. The MCP SDK's client/server split allowed independent transport tests, including negotiation of 2025-11-25 and a real structured tool call. The custom browser UI exposed tool traces and visual confirmations while typed input kept the demo usable without a microphone.
Bedrock Converse supported a compact advisory call with local fallback. Lambda and API Gateway made the two HTTP services inspectable, and IAM/SigV4 checks distinguished signed narrator calls from unsigned requests. DynamoDB conditional operations supported the shared attempt allowance and separate enrollment/session state. Login with Amazon sign-in and sign-out could be verified independently of repair access. Secrets Manager kept runtime configuration out of the frontend. CloudFormation change sets, scoped KMS policies and CloudWatch logs made deployment and operational boundaries reviewable.
TypeScript, Zod and Express made domain/HTTP contracts explicit. pnpm organized shared packages. Docker and GitHub Actions reproduced tests outside the Windows host; observed transaction failures informed regression tests.
Alexa documentation entry points could more clearly distinguish legacy skills, MCP add-ons, category integration and custom simulation. Examples should identify whether they were tested in an official host, a device or only a local client. The MCP SDK would benefit from a protocol-version matrix and complete server/client examples for the exact minimum revision. Browser speech availability varies, and our custom preview cannot validate Alexa host context or certification.
CloudFormation schema validation did not predict the account's Lambda reserved- concurrency quota failure. Account-aware warnings would improve first deployments. Bedrock examples could better show advisory output alongside deterministic business logic. LWA identity must not be mistaken for service authorization or ownership; a worked identity-to-operator-approval-to-redemption example with negative tests would help. IAM simulation was not a substitute for actual authorized/denied calls; KMS conditions needed precise service-path review. DynamoDB backup and restore policies needed explicit treatment so restored state would not undo logout or replenish a model allowance. API Gateway/CloudWatch log configuration and Secrets Manager permissions required deliberate payload/credential minimization.
For TypeScript/Node.js, Zod, Express and pnpm we encountered integration and configuration work rather than a confirmed provider defect. Windows/POSIX differences required Docker tests; CI and code changes still needed source review and real execution evidence.
Alexa+ concepts were understandable, but official end-to-end add-on validation remains incomplete; our documentation review and custom simulator are not proof of partner tooling access. MCP implementation became straightforward after resolving the protocol-version boundary. The browser simulator used ordinary web technology and needed clear labeling plus typed input as a speech fallback.
Bedrock model discovery and the initial Converse invocation were straightforward. Lambda reserved concurrency introduced an account-quota hurdle during CloudFormation deployment. A later reviewed quota/hardening change resolved the reservation issue. API Gateway, IAM and KMS required careful live read-back rather than trusting template validation alone. CloudWatch retention and operational logging were configured explicitly. Secrets Manager configuration stayed private. DynamoDB conditional state and recovery semantics required application-level design, not just table creation.
LWA website login was easier than safe customer mapping: the latter required separate enrollment, private operator authority, customer redemption and actual denial checks. TypeScript/Node.js, Zod, Express and pnpm were familiar building blocks. Workspace-local tooling and Docker addressed Windows host constraints; GitHub Actions supplied Linux coverage. Tests and user-approved hosted operations remained necessary verification steps.
Yes to Alexa+ documentation and the MCP approach for structured service workflows, while reserving judgment on official toolkit/device experience until actually tested. Yes to the MCP TypeScript SDK and the custom web simulator for development: they provide inspectable tool boundaries without requiring commercial APIs.
Yes to Bedrock for bounded advisory text, not pricing or authorization decisions; Lambda/API Gateway for separate HTTP workloads; DynamoDB for conditional state; LWA for identity with independently verified ownership; Secrets Manager for private configuration; and CloudFormation, IAM, KMS and CloudWatch for reviewed deployment, permissions, encryption and logs. We would preserve the distinction between website login and official Alexa service/user authorization. Costs and recovery policies still need workload-specific review; throttles are not a dollar cap.
Yes to TypeScript/Node.js, Zod, Express, pnpm, Docker and GitHub Actions, with pinned dependencies, platform-specific tests and evidence-linked release claims. We have no hands-on conclusion about AgentCore, Kiro Crew, Strands or Alexa devices because those are not implemented/tested parts of this project.