The deck that measures its own trick.

Two families of conductive-fabric strips cross a whole sheet of Velostat, a carbon-loaded polyethylene whose resistance drops under pressure. Every crossing is an analogue cell from 0 to 255. Fabric, not copper: the deck flexes, and copper tape cracks on the first ollie. Edges pinned with Kapton, the centre free to compress. No solder in the standing area — the connection folds to the side of the deck, near the trucks.

The prototype's 8×8 grid came from a handboard demo. Measuring 13,365 frames of real contact in the replay, the foot sweeps 48 cm along the deck by 5 cm across: it slides lengthwise, from pop to catch, not rail to rail. The measured proposal is 40×10 cells of 2 cm over the whole 80.1 × 20.5 cm deck — a sole covers 20 cells and the flick crosses 15 to 25. It costs 8.1 ms of scanning and the second core. Proposed with numbers; not yet decided.

The ESP32 has two cores at 240 MHz and the firmware uses them as two departments. On core 0, the IMU reads at 200 Hz (±16 g, ±2,000 °/s, a 42 Hz filter against asphalt vibration), Madgwick fuses accelerometer and gyroscope into a quaternion, and the matrix is scanned at 60 Hz through two 16-channel multiplexers. On core 1, it all becomes a 114-byte struct with a sequence number and a timestamp, and leaves as a UDP broadcast. Per 5 ms frame the work takes 1.66 ms: 66% is left over.

The IMU sits at the deck's geometric centre, and that enters the maths: 15 cm off centre, a 1,200 °/s kickflip adds 6.71 g of centripetal term to the reading; at the centre, zero. It is the choice that makes the signal interpretable.

MQTT runs over TCP, and TCP guarantees delivery: when a packet is lost, it pauses everything and resends the old one. For kinematic telemetry that is poison — the sample from millisecond 15 arriving at 300 destroys the physics. AXIS speaks UDP from a Wi-Fi network it creates itself, with no router, no internet and no pairing: the same protocol as multiplayer games. Eight bytes of header, no retransmission, latency under 5 ms. A one-byte sequence number reveals what was lost; nobody waits for it.

An event starts at 3 g and is confirmed at 6 g. From there, 1.74 s of acceleration at 50 Hz (decimated from 200) feed a per-axis neural network with 87 inputs, 28 hidden neurons and 9 outputs — the phase-1 flatground tricks — based on Corrêa, Lima, Russomano and Santos (2017). At 70% confidence the deck emits the event: trick name, peak height, landing g, airtime. Classification happens on the microcontroller, before any cloud.

One instant of the Footprint Map, front foot on the left, back foot on the right. The pop is a sudden peak at the tail; the catch is pressure migrating to the centre and nose; the landing is a symmetric distribution with the g peak on the accelerometer at the same instant; the resting position says whether the skater is regular or goofy.

During the session the stream is live: deck → phone or Gappin Node, never through the internet. The ESP32 keeps a copy on its own flash to sync later if the network fails. Session over, the batch goes up as FlatBuffers and becomes one Parquet file per session: ten minutes at 200 Hz and 60 Hz are 7.58 MB raw and about 450 KB columnar, 94% less, queryable in SQL. The dashboard receives a JSON summary, never the raw data.

Every trick becomes a W5H record: who, what, where, when and how — the how is the biomechanical tensor, and in place of why comes WOW, the measured impact. Over that timestamped event stream, the Gappin dashboard does not draw one screen: it draws six projections of the same data. Time is the primary axis and colour encodes data — style polarity, smooth to aggressive, runs cyan to magenta across the whole interface; WOW, 0 to 100 with a half-life, sets the intensity.

Underneath, 55,000 entities in SQLite (skaters, videos, brands, tracks) in a typed graph, the UDP receiver in Rust inside Tauri, and React reading the ring buffer at 20 Hz.