Kilta

A voice-first home repair friend in Alexa+: safety first, manual-grounded steps, any language, clean handoffs when DIY isn’t safe.

Comment

Inspiration

When an appliance breaks, most of us panic. The manual is long gone, we don't know the model number, online advice is generic and sometimes unsafe, and a technician charges $75 to $150 just to diagnose what is often a clogged filter. It's even harder for renters, families more comfortable in another language, and older adults. Existing repair apps live on one person's phone. I wanted a calm, patient repair friend that lives in the home, inside Alexa+, for everyone in the household.

What it does

Say "Hey Alexa, connect me to Kilta." Kilta:

  • Checks for danger first (gas, burning, sparks, water near outlets) before any repair step.
  • Grounds every step in the appliance's own manual, citing the page.
  • Tries the simplest fix first: settings and quick checks before tools or technicians.
  • Enforces safety gates in code: it won't continue until you confirm the appliance is unplugged, and hides hazardous instructions until a safety confirmation is given.
  • Speaks each household member's language, switching mid-repair, while keeping panel labels in English because that's what's printed on the appliance.
  • Remembers repairs across sessions and notices repeat problems.
  • Hands off cleanly: technician summaries and landlord messages written in English, with model, error code, and steps already tried.

How I built it

  • MCP server (Python, official MCP SDK v1) implementing spec 2025-11-25 over Streamable HTTP, with 24 tools, a manual resource, and a prompt. Every tool returns speakable content plus structuredContent for visual cards.
  • Repair engine: a deterministic state machine that owns all safety logic (safety check, re-check, before/after gates, technician-only handoff), so the AI only understands and phrases, and can't skip a safety rule.
  • Agent bridge built with AWS Strands Agents (AWS Builder), acting as the Alexa+ host and calling the MCP server live, with a model fallback chain: Gemini Flash-Lite → Groq → Gemini Flash.
  • Simulator (React, TypeScript, Vite, Tailwind) styled as a smart display, with Web Speech voice input and output in English and Spanish, and a behind-the-scenes panel showing live MCP tool calls.
  • Agent Skill (SKILL.md, agentskills.io format) teaching any agent to use Kilta safely.
  • Tests: 117 server tests, 125 bridge tests, and Playwright end-to-end tests, with CI on GitHub.

Challenges

  • Free-tier AI limits: one model allowed only 20 requests a day, so I built a fallback chain and trimmed each call by about half by sending only the tools that fit the current repair stage.
  • The AI answering for the user: one model asked a safety question and answered it itself. The bridge now cancels a "no danger" answer given in the same turn the question was asked.
  • Safety vs. false alarms: "No, I don't smell burning" contains a danger word, so instead of stopping, Kilta does a quick safety re-check.

What I learned

Safety in AI products belongs in code, not prompts. Live testing with real models catches problems scripted tests can't. And MCP makes a product portable: the same server could plug into real Alexa+ unchanged.

Honest limits

I didn't have Alexa+ Private Preview access, so Kilta runs in a simulated Alexa+ host using the same MCP server. All appliances and manuals are fictional (brand "Kelvara"). The MCP App card and explainer animation are not built yet.

What's next

Real Alexa+ via the MCP Toolkit, MCP Apps visual cards, real manual search, recall checks, more languages, and opt-in anonymous repair insights for brands and shoppers.

Built With

Share this project:

Updates

Submission history