BASIC coaches a battery installer through Meta Ray-Ban glasses. Every two seconds the glasses camera sends a photo to Claude, which checks it against the current step of the job. A voice in the glasses confirms the step, corrects it or gives the next one. A master electrician watches every crew from one dashboard and can talk into anyone's glasses by typing.

Built for Base Power's install crews at the Base Power x AITX hackathon, Austin, September 2026.

Quick start

git clone https://github.com/sisiphamus/BASIC.git && cd BASIC
npm install && npm run build
cp env.example .env     # MODEL_PROVIDER=claude-cli
npm start               # dashboard at http://localhost:3000

No model yet? MODEL_PROVIDER=mock npm start runs a fake model for rehearsing, and npm run simulate -- --crews 3 --loop fills the dashboard with simulated crews.

Architecture

flowchart LR
    G["Meta Ray-Ban<br/>glasses"] -- "photo every 2 s" --> A["Android app"]
    A -- "1080×1440 JPEG" --> S["Node server"]
    S -- "photo + step + check" --> C["Claude Sonnet"]
    C -- "verdict as JSON" --> M["State machine"]
    M -- "next line" --> A
    A -- "voice over Bluetooth call audio" --> G
    S -- "live over WebSocket" --> D["Supervisor dashboard"]

Loading

• The model sees, the code decides. Claude returns pass, fail or unclear with its evidence and a confidence score. A deterministic state machine only advances the job at 60% confidence or above.
• It always judges the newest photo. One model call runs per crew at a time, newer photos replace older ones in the queue, and answers for a step the tech has already left are thrown away.
• Hands are the interface. Each step ends with a gesture (point at the studs, a hand on the part, a thumbs up) that tells Claude the tech is ready to be checked.
• Jobs are YAML playbooks. They're validated on load and reload on save, so a step can be rewritten while someone is wearing the glasses.
• The voice works while the camera streams. The glasses mute media audio during a stream, so speech goes over the Bluetooth phone-call channel.
• Everything is logged: each photo with its quality stats, the prompt, Claude's raw answer, cost and the decision. npm run console streams it live.

Measured on real glasses: about 5 seconds and $0.013 per check with Claude Sonnet on low effort.

Tech stack

Layer	Tools
Glasses	Meta Wearables Device Access Toolkit 1.0 on Ray-Ban Meta
Phone app	Kotlin, Jetpack Compose, Android TextToSpeech, OkHttp
Vision	Claude Sonnet via the Claude Code CLI, or the Google Gemini API
Server	Node.js 22, Express 5, WebSockets
Jobs	YAML validated with Zod
Dashboard	React 19, Vite, Tailwind CSS 4
Tests	node:test (93 tests), Playwright

Reproduce the demo

You need Node.js 22 and the Claude Code CLI logged in, or a Gemini API key. For the glasses you also need an Android phone, Ray-Ban Meta glasses, JDK 17 and the Android SDK.

1. Configure by copying env.example to .env. Claude needs no key in the file, because the server calls the logged-in claude command.

MODEL_PROVIDER=claude-cli      # claude-cli, gemini or mock
# BA_CLAUDE_MODEL=sonnet       # BA_CLAUDE_EFFORT=low
GEMINI_API_KEY=                # only for MODEL_PROVIDER=gemini
# PORT=3000                    # HTTPS_PORT=3443, DATA_DIR=./data

2. Start the server with npm start. The last startup line names the model in use.

3. Install the glasses app. Turn on Developer Mode for the glasses in the Meta AI app, then:

cd android && ./gradlew assembleDebug
adb install -r app/build/outputs/apk/debug/app-debug.apk
adb reverse tcp:3000 tcp:3000     # phone on USB reaches the laptop

Open BASIC, set the server to http://localhost:3000, tap Connect glasses once, pick Base battery install (visit 2) and press Start.

4. Run the 12 steps in the demo script. No glasses? Turn on "Test without glasses" and Meta's MockDeviceKit supplies simulated ones.

Data

Nothing was trained. Every step is judged against a plain-English check written in the playbook.

• Sample photos come from Wikimedia Commons (CC0, public domain and CC licenses), each credited in demo-frames/CREDITS.md.
• Synthetic: the installer app screens, the serial label, and the customer and job details are made up for the demo.
• Install sequence is based on a Base Power installer's description at the hackathon. It is a teaching demo, not Base's official procedure.
• Real glasses photos from our runs stay local in data/ and are not in the repo.

Known limitations

• About 5 seconds per check: fine for step-by-step work, too slow to catch a fast slip.
• It judges single photos, so it can't measure torque or voltage.
• The QR code on the real battery is too small to read at 1080×1440, so the scan step only confirms the label is in view.
• The glasses drop the session when taken off, and reconnecting sometimes takes an app restart.
• There's no login, so run it on a trusted network. Tested only on a Galaxy S24 with Ray-Ban Meta.

Next steps

• Full-resolution stills on scan steps, to read the serial and match it to the job.
• Plug into Base's real installer app and job records.
• An iOS app, plus login and per-crew access.
• Turn passed steps into sign-off records a master electrician approves.

---

Field guide · Demo script · Image credits