Symbionic EMG Studio is a desktop workstation for the muscle signals that drive a bionic arm, built so a clinician can calibrate the arm for a patient and see exactly how well it reads them.
The Symbionic sensor board sends six channels of EMG, and the grip the arm has decoded, as text lines over USB serial, about 250 a second. The app draws all six live on one time axis with the decoded grip beneath, records sessions under a subject ID with markers, and reduces each grip to a signature: every channel's RMS while the grip is held, estimated from the median of squares so a single saturation spike cannot skew it.
Around that sit a dashboard, a searchable session log, channel detail with a power spectrum, grip analysis, a serial monitor that explains every line it refuses, and a second window that walks a patient through electrode checks and a guided calibration. It ships as two editions from one codebase: the standard edition connects to the board, and the demo, the download on this page, has a simulated board built in, so every screen works with no hardware.
In 2023 I built a prototype of this during an internship at Symbionic, and it could only chart a saved file. I rebuilt it as the tool a clinician actually needs.
After a guided calibration the clinician sees, for each grip, how often the arm decoded the grip it was asked for, how distinct the grips are from one another and whether every electrode was clean at rest. Then one sentence: ready to use, worth another pass on one grip, or check the electrodes. The patient's reaction time is left out of the score, so a slow start is never counted as a misread.
Samples never pass through React. They land in a two-minute buffer, and uPlot draws each lane from a zero-copy view of it, running slightly ahead of the data between the 30 Hz batches so the trace scrolls smoothly instead of stepping. Each lane scales to its own signal and clips a saturation spike at its edge, so one bad sample cannot flatten the trace. Drag zooms, scroll pans, and M drops a marker into the recording.
File, New patient session opens a second window built for the appointment: a patient ID that is never a name, an electrode check on each channel's noise floor with the arm relaxed, then each grip prompted in turn with a countdown while the whole run records. Both windows read one broadcast stream, and a window asks for the current state the moment it opens, so it can never miss what was sent before it was listening.
The demo's board is modelled on a real recording from the actual board: 9,750 samples at 250 Hz, each grip driving the six channels at the RMS measured there. It writes text lines into serialport's mock port, so the app reads them through the same parser, batching and recorder as a real COM port. During calibration it follows the prompts with a human reaction time and the odd misread, so results look like a real fitting rather than a flat 100%. Its fault mode corrupts lines on purpose, and the serial monitor shows each refused line with the reason.
Every session is a plain CSV that opens in Excel, MATLAB or pandas, plus a JSON sidecar with the subject, markers, a summary and, for a calibration, its result. A marker that looks like a spreadsheet formula is written as text, so opening a file in Excel cannot run it. The main process only touches files it listed or the user picked, and a deleted session goes to the Recycle Bin rather than disappearing.
One channel with its 100 ms RMS envelope, a Welch power spectrum with mains hum marked, and the noise floor, contraction level, signal-to-noise and median frequency. The 20 Hz high-pass and the 50 or 60 Hz notch are display-only, so what is recorded is always what the board sent.
A prototype in 2023, then a rebuild that kept what was true: the board's wire format, a real recording and a clinician's job. The core and its tests came first, the screens after.
A labelled recording from the Symbionic sensor board: rest, then each of three grips, three times over. It later became the simulator's source of truth and the test the simulator has to pass.
An Electron app that opened a saved recording and charted it. The live view over serial, settings and calibration were still unfinished.
The wire format, a simulator checked against the real recording, the recording format and the settings, as one shared module with its own tests before any screen existed.
The board, the simulator and file playback behind one typed IPC, with context isolation and the sandbox on.
Live lanes, the decoded grip, recording and playback, then the dashboard, the session log, channel detail, grip analysis and the patient window.
A per-user Windows installer for each edition, a release built from a version tag, and an end-to-end run of the real app across every screen.
A native Windows app with real menus, dialogs and the Recycle Bin. Context isolation and the sandbox are on, so windows never see Node, the file system or the port.
One strict shared module for the wire format, simulator, analysis and session format, built once for the main process, the preload and both windows. The edition is fixed at build time.
React draws the screens and zustand holds what the UI needs to remember. The signal itself never passes through either.
A canvas chart small and fast enough to redraw six 250 Hz lanes at 60 fps from views of a typed buffer, with no copying.
N-API prebuilds, so there is no native compile step, and a mock port that lets the simulator use the exact path a real board does.
Settings and session sidecars are checked on every read, so a hand-edited or damaged file falls back to safe values instead of crashing, and settings are written atomically.
25 tests on the parser, the simulator against the real recording, the summaries, calibration agreement, spike resistance and the filters.
27 checks that drive the real app, or the installed exe, through every screen and the patient window against a throwaway profile.
Tests on every pull request and every push to main, and the installer built on a version tag and attached to a draft release, never from a laptop.
A per-user installer that needs no admin rights, from two configs: the demo and the standard edition share everything but their data source.
The demo ships with a simulated board and example sessions and connects at launch, so there is nothing to ask a visitor first.
Bundled rather than fetched, so the app looks the same offline. Inter for everything read, JetBrains Mono for every figure.
The standard edition is the same app with the board in place of the simulator, and it already builds from this codebase. What stands between it and a clinic is specific.
The wire format comes from the 2023 recording. It is confirmed against the board's current firmware and baud rate before the standard edition ships.
The installer is not code-signed yet, so Windows SmartScreen asks once. Signing removes that question.
Updates are off until there is a signed feed to take them from.
AlexanderSmith
Symbionic EMG Studio
