★ 4th Place · BeaverHacks 2026 — NVIDIA Track

LINX

Local-first AI that watches over people living with Alzheimer's — before they know they need help.

Instead of waiting for someone to ask for help — which requires them to realize they need it — Linx proactively monitors the environment with low-cost sensors and infers intent in real time, fully offline. Built in 24 hours with a team of four; placed 4th in the NVIDIA track at BeaverHacks (Oregon State University).

Timeline2026 · 24h Build
RoleAI & Systems Engineer
TeamBeaverHacks — Team of 4
Outcome4th · NVIDIA Track

The Challenge

People living with Alzheimer's often can't recognize when they need help — so reactive assistance, which waits for a request, fails exactly when it matters. Memory-care facilities are understaffed, and existing monitoring tools are cloud-dependent, sending intimate behavioral data off-device. The challenge: detect risk proactively, in real time, for a population that deserves both safety and privacy.

The Solution

Linx runs entirely on a ~$150 hardware stack — a Raspberry Pi with a USB camera, ultrasonic sensor, and 4-mic array. Sensor data flows into a rolling memory buffer reasoned over by NVIDIA's Nemotron 3 model running locally via Ollama. A parallel deterministic rule engine catches high-risk patterns instantly. Nothing leaves the device — privacy was a day-one constraint, not a feature.

01

Sensor Fusion & Intent Inference

Camera, ultrasonic, and 4-mic-array signals are structured into a fixed-size rolling memory buffer — short enough to keep prompts lean, long enough to preserve context — then reasoned over by Nemotron 3 to infer what the person is doing and what they may need next.

  • Multi-sensor event loop on Raspberry Pi
  • Rolling deque memory
  • Real-time intent inference
02

Nemoclaw Safety Rule Engine

A deterministic rule engine runs in parallel with the model, evaluating known high-risk patterns — stove left unattended, extended inactivity, door left open — before inference even completes. Rules can escalate the model's risk assessment but never lower it, so safety-critical events are caught instantly and reliably.

  • Deterministic checks beat model latency
  • Rules raise risk, never suppress it
  • Reliable critical-scenario coverage
03

Privacy-First, Fully Offline

The entire system runs offline via Ollama — no API calls, no cloud latency, no sensor data leaving the home. Beyond safety alerts, Linx can notify emergency contacts, generate timestamped behavior reports for medical staff, and flag when nursing support is needed.

  • Zero data egress by design
  • Emergency-contact notifications
  • Timestamped reports for care staff

Hardware

Raspberry Pi · DFRobot USB Camera · Ultrasonic Sensor · ReSpeaker 4-Mic Array

AI & Reasoning

Nemotron 3 · Ollama · Nemoclaw Rule Engine

Backend

Python · MCP Microservices · Rolling Memory Buffer

Principles

Offline-First · Privacy by Design · Edge AI

4th NVIDIA track · BeaverHacks
$150 Total hardware cost
100% Offline — zero data egress
6 Independent MCP services
01

Determinism Where It Counts

For safety-critical detection, a deterministic rule layer that runs before — and can only escalate — model inference proved more trustworthy than relying on the LLM alone. Latency and reliability both improved.

02

Microservices Enable Swapping

Splitting the backend into isolated MCP services let us swap a quantized model for constrained hardware or add a new sensor without touching the reasoning layer — invaluable under hackathon time pressure.

03

Privacy as a Constraint, Not a Feature

Designing for fully offline operation from day one shaped every architecture decision. For a vulnerable population, keeping data on-device wasn't a nice-to-have — it was the foundation.

04

Build for a Real Problem

Having volunteered at an assisted-living center, the team built around lived experience of how disorienting daily life can be with Alzheimer's. A concrete human need kept scope and priorities honest.

Next project Healthcare ML