Kompic̄

Project Overview

In Active Development

Kompic̄ is a fully open-source, wrist-mounted sensor platform built by an ML scientist for on-device inference. It pairs a sensor suite denser than any open-source watch with an ESP32-S3 compute platform powerful enough to run real neural networks on that data — sleep staging, sound classification, sensor fusion — all locally, all offline, all inspectable.

The name is Croatian street slang: komp ("computer") plus the diminutive -ic̄ — a "little computer," and the word you'd only use for an actual friend, not a stranger. Pronounced /ˈkompitʃ/ (kompich).

Author

Hardware, firmware, case and webpage design by Ivan Porupski, 2026.

Philosophy

• Open source — hardware (KiCad), firmware, ML models, and case design (Onshape). Everything inspectable, modifiable, forkable.
• Offline first — works fully standalone. No phone required, no cloud dependency, no subscription. BLE is a capability, not a requirement.
• Raw data access — PPG waveforms, ECG traces, IMU streams, mic audio, environmental readings. Not just processed metrics.
• ML-native — the ESP32-S3's compute surplus exists specifically to run on-device inference. The sensor suite is designed to feed models, not just display numbers.

Key Specs

• ESP32-S3 dual-core Xtensa LX7 @ 240 MHz, 8 MB PSRAM, 16 MB flash
• MicroSD storage (128 GB+), BLE 5.x
• 2.06" rectangular AMOLED, 45.6 × 37.3 mm glass
• Two 1.0 mm 4-layer PCBs: main board + skin-facing daughter board
• Two-piece titanium case (SLM Ti-6Al-4V), ~48 × 40 × 12 mm
• ~380 mAh LiPo, USB-C charging

Sensor Suite

Circuit + PCB

[Circuit + PCB Guide →]

Two 1.0 mm 4-layer PCBs designed in KiCad: a main top board carrying the ESP32-S3, display interface, power architecture, and most sensors, and a skin-facing daughter board carrying the MAX30101, TMP117, and haptic driver. Schematic capture is complete; routing is underway.

• Full power tree: BQ25619 charger → SYS → TPS62840 3.3 V buck → 1.8 V LDO, plus FW-gated 5 V boost
• Display: CO5300 AMOLED driver over QSPI; CST9217 capacitive touch
• Two I2C buses, controlled-impedance RF traces for BLE and GPS antennas
• KiCad schematic, PCB layout, and JLCPCB production files (coming soon)

Fabrication + Assembly

[Fabrication Guide →]

Planned

Boards are fabricated and partially assembled by JLCPCB — 0402 passives placed, with baked solder paste applied to the remaining IC and connector pads. ICs and connectors are then hand-placed and hot-plate reflowed by the builder. Full production .zip files and a picking list will live in the repo.

Manual Assembly

[Assembly Guide →]

Planned

Step-by-step hand-assembly walkthrough: hot-plate reflow profile, stacking the display on the main PCB, seating the battery, mating the daughter board, wiring the 12-conductor silicone interconnect, and closing the two-piece case. MSL3 bake notes for the ESP32 module and GPS included.

3D Models & Printing

[3D Printing Guide →]

Planned

Single-piece case plus a separate crown. The body is one 3D-printed titanium shell — no front-plate/back-body split and no assembly screws. Everything loads in from the open display side and is captured by the body's internal features; the display, attached on last with a 1 mm double-sided tape, closes the case. The crown is the only separate machined part — ordered oversized and finished by hand.

Prototypes are modeled in Onshape and printed in plastic on a Bambu Lab A1 mini for fit testing; the production case is SLM titanium (Ti-6Al-4V). Onshape source files and print settings (coming soon).

• Fastener-free internal retention: side prongs hold the main PCB at height, a recess seats the USB-C jack, and a dedicated pocket holds the LRA
• Optical/RF windows (BLE, GPS, status LED, flashlight, ambient light): epoxy-filled through-holes — clear where light passes, opaque where it's only structural
• Mic + BME688 air port: a PTFE sticker inside, acoustic/gas permeable and water resistant
• Open-source case files (coming soon)

Firmware

[Firmware Guide →]

Planned

Dual-core firmware: one core handles hardware drivers and sensor acquisition, the other drives the LVGL UI on a DMA-based, double-buffered AMOLED pipeline in PSRAM. On-device ML runs on the ESP32-S3 for sleep staging, snore/sound classification, and sensor fusion.

• Display pipeline — DMA double-buffer architecture
• Sensor drivers — validated raw-data quality across the full suite
• Sleep tracking — HR + temp + actigraphy → ML inference → smart wake
• Navigation — GPS + compass fusion → breadcrumb trail
• Audio — mic capture → snore detection → voice commands

Real-Life Application & Data Analysis

Planned

The primary use case is sleep and recovery: three-signal fusion (HR/HRV, skin temp, actigraphy) for sleep-stage estimation, gentle haptic smart-wake during light sleep, and overnight environment scoring from the air-quality, light, and noise sensors. This section will collect real captured datasets, analysis notebooks, and notes on how the on-device models perform against the raw signals.

• Sleep staging — captured nights vs. model output
• HRV / recovery trends derived from raw PPG
• Environment logs — temperature, humidity, VOC, noise, light
• Navigation tracks — GPS + barometric altitude

References, Blog & Misc

Planned

Build logs, design-decision write-ups, datasheets, and the open-source landscape this project sits in (ZSWatch, Bangle.js, Watchy, PineTime and others) will be collected here.

1. [1] Project scope & market position — internal doc, v1.0
2. [2] ZSWatch — zswatch.dev
3. [3] Bangle.js 2 — banglejs.com
4. [4] Watchy — watchy.sqfmi.com