Inspiration

Everyone does it. You're deep in a hackathon, a library, or an office, and you walk away from your laptop for "just a minute." Your screen lock doesn't stop anyone from plugging a USB into your machine, and when you come back you have no idea whether anyone touched it. We wanted a laptop that guards itself while you're gone and tells you the truth when you get back: a bodyguard and a detective in one.

What it does

You press Start Protection and walk away. From then on, LOCKI watches your workstation from two independent angles:

  • Physical: an NVIDIA Jetson with its own camera watches your desk zone for a person.
  • Digital: your Mac watches for new USB devices and app launches in real time.

The moment something suspicious happens, like an unknown USB plugged in while you're away, LOCKI speaks up out loud and warns the intruder off. That warning plays from a local voice cache, so it never waits on the cloud or an AI model.

Every event flows into SpacetimeDB, which acts as the single source of truth. It decides what counts as an incident, correlates camera and USB evidence inside the same time window, raises the severity, and keeps a tamper-evident, ordered audit trail.

When you come back, you just ask "LOCKI, what happened?" An NVIDIA Nemotron agent investigates the incident. It uses read-only tools to pull the events, the USB device, the camera observations and the baseline from when you armed it, and then answers out loud. It explains what is known and what isn't, and cites the evidence behind every claim.

How we built it

LOCKI runs across four machines that work as one system:

  • M4 MacBook (the guarded workstation): a Python/FastAPI backend with native macOS USB (IOKit) and app-launch watchers, a priority speech queue with local Kokoro TTS that follows whatever speaker is connected, and a React + Vite command center with Protection, Investigation and History views.
  • SpacetimeDB (maincloud): a TypeScript module that owns enrollment, roles, armed sessions, event correlation, incidents and lock requests. Every device has its own identity and only sees caller-scoped views of its own data.
  • NVIDIA Jetson: our edge runtime runs MobileNet-SSD person detection on the camera, with a durable outbox so no events are lost when the network drops. It publishes only sanitized observations: no images and no faces.
  • M5 MacBook Pro: runs nemotron-3.5-bolt locally through Ollama behind an authenticated investigator API. The M4 reaches it through an SSH tunnel over Tailscale, so the model and the evidence never touch the public internet.

We built on the voice engine from LOCKI, our earlier voice assistant project. This weekend we built everything security: the watchers, the SpacetimeDB module, the Jetson runtime, the Nemotron investigator and the command center. A one-click Desktop launcher checks every component for real (backend, database, Jetson heartbeat, Nemotron answering, volume) and gets the demo to READY in about 1 second.

Privacy and safety by design

  • No face recognition. The camera reports presence, not identity.
  • The owner is the only authority. Voice can't arm or disarm, and arming always needs the owner's explicit, authenticated action.
  • Evidence needs consent. Nemotron can only read an incident after the owner grants access to that specific incident, and the grant is revoked when protection stops.
  • Locks are dry runs unless the owner opts in, and USB serials leave the Mac only as salted hashes.
  • Local only. The backend refuses any request that isn't from the machine itself.

Challenges we ran into

  • Making four machines agree. A Jetson, two Macs and a cloud database all had to share one truth. Putting every decision in SpacetimeDB reducers, instead of in each device, is what finally made incidents consistent everywhere.
  • Speed vs. intelligence. An LLM investigation takes seconds, but a warning has to be instant. We split the two: the warning plays from cache in about 335 ms from event to audio, and Nemotron enriches the incident afterwards without ever blocking the alarm.
  • Networks. The venue network blocked direct peer-to-peer connections, so the M4→M5 link runs through a Tailscale relay with a self-healing tunnel supervisor that reconnects in seconds.
  • Security of our own app. We found that a malicious website could reach our local API through DNS rebinding, so we added a loopback-only host and origin guard to the backend.

Accomplishments that we're proud of

  • A real end-to-end incident: a person at the desk plus an unknown USB, a spoken warning, a correlated incident in SpacetimeDB, and a Nemotron investigation that ran 7 read-only tools in about 12 seconds.
  • Nemotron working as a real tool-using investigator, not a chatbot, and running entirely on our own hardware.
  • 2,500+ automated tests across the backend, investigator service and dashboard.
  • An honest system: synthetic test events are always labeled, and LOCKI never claims a lock or a finding it didn't make.

What we learned

Security products are mostly about trust boundaries. We learned to keep the AI advisory and the database authoritative, to make consent explicit and scoped, and to treat every device as untrusted until it's enrolled. We also learned to make the fast path never depend on the smart path.

What's next for LOCKI AIRLOCK

  • Real opt-in screen locking and phone notifications for the owner
  • Accuracy trials with consenting volunteers to tune false alarms
  • Multi-desk support for labs, coworking spaces and offices
  • A small all-in-one LOCKI desk unit built on the Jetson

Built With

Share this project:

Updates

Submission history