- Start from open hardware
The idea was an EEG headband you could sleep in, recording night after night at home so a clinician has objective data rather than a questionnaire. Rather than design an analog front end from scratch, the prototype started from OpenBCI's Ganglion (a four-channel biosensing board published as open hardware) and reworked it: re-captured in KiCad 8, re-laid-out, and re-sourced part by part.
- What “rework” meant
The project's own files, not OpenBCI's: 140 footprints on a 61 mm four-layer octagon, a BOM of 41 lines, and a report of every symbol rescued from the old libraries. The layer stack on this page is rendered straight from that .kicad_pcb.
- Sourcing
Four parts were obsolete or overpriced and were swapped with checked equivalents: a Kemet 1 µF for the Murata, a Samsung 47 µF, a 100 V X7R upgrade for the 0.1 µF decouplers, an onsemi supervisor variant. Two parts could not be sourced at all: the Simblee BLE module (discontinued) and the Seeed microSD socket. The DigiKey order for everything else came to $64.28.
- The signal chain
Four AD8237 instrumentation amplifiers feed an MCP3912 four-channel 24-bit delta-sigma front end; TVS arrays protect the electrode inputs, DIP switches pick between the header pins and the input pads, and a DAC-driven reference sets the bias. The schematic tour walks each block and lights up the matching parts on the board.
- Recording HELLOworld
A first recording with a lab Cognionics headset (20 EEG electrodes in the 10-20 layout at 500 Hz), about a minute long, exported as BrainVision. The header's impedance table tells the story: Cz, T3, C4 and A2 had poor contact, and A2 is almost entirely 60 Hz mains. It ships here at 250 Hz as 16-bit integers, and everything you see is computed from it in the browser.
- The DSP notebook, finished
The notebook loaded the recording with MNE, computed Welch coherence between channel pairs, and had a filter-and-downsample cell that was never finished. Each step is ported to TypeScript here (FIR lowpass and decimation, the completed cell; a 60 Hz notch; common-average referencing; Welch cross-spectra) and the ports are tested against SciPy on a slice of the recording. Coherence between electrode pairs (how much two regions oscillate together at each frequency) was the feature the venture wanted to track across nights.
- The business case
The Basement's launch program at UC San Diego framed it as a diagnostic aid: overnight EEG plus a model to help psychiatrists tell bipolar disorder from depression, sold as a headset and a monitoring subscription. A Business Model Canvas, a Value Proposition Canvas and an Empathy Map from November 2024 close the page; the venture stopped at this prototype.