For Investors
Engineering
Edge-first embedded system: an ESP32-S3-WROOM-1 runs a MobileNetV3-Small INT8 dog-pose classifier and an 80Hz storm-audio classifier on-device at 5fps, driving a single-flight damped auger through a DRV8833. A companion mobile app pairs over BLE for setup and reads training progress; a thin cloud control plane handles OTA model/firmware, anonymized session telemetry, and Stripe subscription. Safety-critical inference never gates on the cloud, and no raw frames or audio ever leave the unit — the on-device-only design is the deliberate BIPA / two-party-consent mitigation.
Front-to-back section through the center axis of the tower showing the gravity-fed treat path from the top-loading hopper, down the damped Delrin auger, through the acoustic-foam chamber, out an external chute into the floor-reachable silicone tray. Electronics stack low-front (camera, mic array, IR LEDs) and low-rear (main PCB, USB-C power board), with the steel weighted plate at the very base for tip resistance.
Dispense tray is drawn external to the sealed shell, below the base outline — the dog reaches into open air, never into the housing.
Chip selection
ESP32-S3-WROOM-1-N16R8 chosen as a single-chip compute + radio that runs the classifier on-device without an external NPU or a second radio/cert path.
Sensor & camera architecture
A low-forward camera reads body language at the dog's own height, a beamformed MEMS mic array detects storms independently, and IR assist handles dark 3am storms — none of it a startle source.
- OV5640 5MP camera
- ICS-40180 PDM MEMS mic x4
- 940nm IR LED x8
ML inference pipeline
All safety-critical inference is on-device; the device ships dispense-DISABLED (log-only) until the classifier is validated, and the cloud never gates a dispense decision.
Power & battery
USB-C mains, deliberately no battery in the dispense path — a dead battery at 3am during the exact event the product exists for is an unacceptable failure mode.
Firmware architecture
FreeRTOS with the safety-critical inference loop isolated from network jitter; the single most important property is that the calm-confidence floor cannot be lowered from the app.
Manufacturing process
- Single-cavity injection mold for the tower + base shells ($35K, replaces a $105K three-SKU tool)
- SMT PCBA line + AOI (4-layer proto -> 2-layer + jumper at scale after RF validation)
- Auger + acoustic-chamber sub-assembly (Delrin flight + Basotect/MLV)
- Functional end-of-line test: <40 dB dispense, jam-recovery, fail-closed gate
Architecture summary
Edge-first embedded system: an ESP32-S3-WROOM-1 runs a MobileNetV3-Small INT8 dog-pose classifier and an 80Hz storm-audio classifier on-device at 5fps, driving a single-flight damped auger through a DRV8833. A companion mobile app pairs over BLE for setup and reads training progress; a thin cloud control plane handles OTA model/firmware, anonymized session telemetry, and Stripe subscription. Safety-critical inference never gates on the cloud, and no raw frames or audio ever leave the unit — the on-device-only design is the deliberate BIPA / two-party-consent mitigation.
Risk register
Risk register
| Risk | Severity | Likelihood | Mitigation | Confidence |
|---|---|---|---|---|
| On-device classifier misclassifies during an anxiety event and rewards a fear state (iatrogenic) | Critical | Medium | Ship dispense-DISABLED (log-only) until >95% calm/anxious accuracy + <2% false-positive fear-reward on held-out data; 0.95 fail-closed floor is firmware-locked; documented audio-only pivot if unmet | Modeled |
| OV5640 camera module goes EOL mid-production or ESP32-S3 allocation slips | Medium | Medium | Qualify a pin-compatible second camera source before the 1K run; order ESP32-S3 ahead with buffer stock | Sourced |
| <40 dB auger noise target not met across variable treat sizes | High | Medium | Basotect+MLV acoustic chamber, single-flight fixed-rotation auger, current-sense stall recovery; validated in EOL test BEFORE tooling commit | Modeled |
| FCC intentional-radiator + UL cert slips past the launch window | Medium | Low | FCC pre-scan at design freeze; cert runs parallel to the mold lead time; budget reserved in the capital ask | Modeled |
Dimensions (to scale)
Overall dimensions by model
| Model | Height (mm) | Width (mm) | Depth (mm) | Weight (g) | Note |
|---|---|---|---|---|---|
| Medium | 165 | 130 | 190 | 1050 | Depth includes 45mm tray extension; below shoulder height for 20-60lb dogs; weighted base ~200g of the 1050g total |
Component dimensions (to scale)
| Component | Value (mm) | Tolerance | Note |
|---|---|---|---|
| Camera bore | 14 | ±0.15mm | OV5640 module seat, low-forward face |
| Mic array span | 50 | ±0.5mm | 4-mic diamond beamforming geometry |
| Tray recess depth | 8 | ±0.3mm | prevents treat roll-off, 6° forward tilt |
| Tray external extension | 45 | ±1mm | forward of base shell, floor-reachable |
| Hopper capacity height | 130 | ±1mm | ≈120 standard 12-15mm training treats |
| Acoustic chamber wall | 12 | ±0.5mm | Basotect + MLV, drives <40dB target |
| Base weighted plate mass | 200 | ±5g | steel insert, tip resistance (value in grams, not mm) |