Myelix

Resilience is hard to measure because it's everywhere and nowhere at once — in how you type when you're stressed, how you walk when you haven't slept, how steady your touch is on the day before a migraine. Myelix is the system that watches those signals with your permission and tells you what your body is trying to say.

What it is

Myelix is a privacy-first resilience monitoring system. A mobile app collects passive behavioral signals — typing cadence, gait, touch patterns — establishes a 30-day personal baseline, and detects drift from that baseline using on-device machine learning. Raw biometric data never leaves the device.

The system is split into two coordinated halves:

• myelix-community (public) — the .NET 10 MAUI mobile app for Android, iOS, and Windows. Passive sensor abstraction layer, local baseline engine, ONNX inference runtime, encrypted SQLite storage, and a Plugin SDK for researchers.
• myelix-core (proprietary) — the server-side research API, drift scoring engine, and federated learning pipeline that ingests anonymized aggregates only — never the underlying raw signals.

The two halves communicate via versioned NuGet contracts, so the public client and the private research infrastructure evolve independently without coupling.

What it does well

Passive, not performative. You don't open Myelix to log a feeling. It runs in the background, observing the signals you already produce just by using your phone, and learns the shape of your normal.

Ghost Mode baseline (30 days). The first month is observation only. No alerts, no nudges — Myelix is figuring out what your normal looks like, not anyone else's. After the baseline is established, the inference engine has something to compare against.

Strictly on-device inference. ONNX Runtime + TCN model running locally. Your typing cadence and gait data never go to a server. Even drift scores stay on-device by default.

Federated learning for the model, not for your data. When you opt into population research, Myelix sends anonymized aggregates through the federated learning pipeline. The model improves; your raw signals never leave your phone.

Extensible via Plugin SDK. Researchers can build plugins against the public sensor abstraction layer without needing access to the proprietary scoring engine — a clean separation between collecting signals and interpreting them.

Who it's for

People managing chronic conditions where early-warning signals matter — migraine, MS, fibromyalgia, chronic fatigue, mental-health flare-ups.

Athletes and performance-focused users who already track recovery and want a passive layer beneath the explicit one.

Researchers studying behavioral biometrics, drift detection, or federated learning patterns who want a real client with a real sensor pipeline to build against.

Privacy-first users who refuse to install another app that ships their data to a cloud they don't control.

What it looks like in practice

A user installs the Myelix mobile app on Day 0. For thirty days, it observes — no alerts, no notifications, just learning. On Day 35 the user wakes up feeling fine, but Myelix has noticed that her typing rhythm has been drifting for the past two days. She gets a low-key notification: "Baseline drift detected. Pattern resembles your previous fatigue episodes." She drinks more water and takes the afternoon off. Two days later, the drift normalizes.

What it's not

It's not a medical device. Myelix doesn't diagnose. It detects drift, surfaces it, and lets the user act on it.

It's not a social product. There is no leaderboard, no comparison to other users, no "you're doing better than 60% of people" framing.

It's not surveillance. The privacy boundary is the device. Nothing crosses it without an explicit opt-in to the anonymized aggregates pipeline.

Status

Active development. Public client (myelix-community) on .NET 10 MAUI for Android/iOS/Windows. Private research half (myelix-core) on ASP.NET Core 10 with a Python 3.12 + PyTorch training pipeline exporting ONNX models. Self-hosted CI on Proxmox runners. Plugin SDK published as versioned NuGet from the public repo.