Pawvlov NLT Labs NLT Labs

For Investors

Engineering

ESP32-S3 dual-core MCU drives a 4-MEMS PDM microphone array (storm detection + classification at 80 Hz) plus an OV5640 120-degree wide camera running on-device MobileNet v3 INT8 inference for dog-pose classification (calm | alert | anxious | absent), plus a near-silent N20 12V gear motor + Delrin helical auger for treat delivery (<40 dB at 1 m). The companion mobile app pairs over BLE and syncs telemetry over WiFi; a thin cloud control plane (FastAPI + Postgres on Render, R2 object storage for 30-sec session clips) handles cloud-fallback inference, OTA firmware, Stripe billing, and session-history aggregation. Mains-powered via USB-C - no battery in the dispense path so brownouts cannot mis-time a treat during a session.

Vertical front-elevation cross-section through the Pawvlov M chassis (140x140x165 mm). Top to bottom: snap-close hopper chamber with 35-degree gravity funnel feeding the N20 gear motor and Delrin helical auger inside an acoustic-foam-lined chamber targeting less than 40 dB at 1 m; OV5640 camera and 4x ICS-40180 MEMS mic array on the front face at dog-eye level; 8x 940 nm IR LEDs in the lower body for low-light camera assist (invisible to canine vision); 4-layer ESP32-S3 main PCB in the lower bay; 40mm I2S speaker at lower-front; SK6812 RGBW LED ring at the base perimeter. Sage-green TPE treat tray shown EXTERNAL — dog must approach the device to eat, mechanically reinforcing approach behavior as part of the DACVB counter-conditioning protocol. S/L variants share all PCB and electronics; enclosure height and hopper volume scale per-size.

Internal cross-section (SVG)
Pawvlov — Internal cross-section (SVG)Labeled engineering cross-section of Pawvlov showing internal components, dimensions, and wiring.Section A-A · front elevation cutH 165 mmW 140 mmTREAT HOPPER CHAMBERABS food-safe · 180 treats (M) · anti-bridge ribs · 35 deg gravity funnel35 DEG GRAVITY FUNNELDelrin POM · single-treat throat 13 mm (M) · anti-bridge geometryACOUSTIC FOAM CHAMBERClosed-cell melamine · 12 mm wall · less than 40 dB targetN20 12V GEAR MOTOR + DELRIN AUGER100 RPM · Delrin helical · DRV8833 driver · 38 dB at 1 mOV5640 · 120 DEG LENS5 MP · downsampled 224x224 at 5 fps · 940 nm IR-cut removed · FFC to PCB4x ICS-40180 PDM MEMS65 dB SNR · 12 mm spacing · 80 Hz storm detect · beam-formingESP32-S3-WROOM-1-N16R8 · MAIN PCB4L 100x70 mm · dual-core 240 MHz · WiFi+BLE · MobileNet v3 INT8 · DRV8833 · PMIC40 MM 5W I2S SPEAKERMAX98357A Class D amp · 200 Hz floor · calm-audio reinforcement8x 940 NM IR LED ARRAYVishay TSAL6400 · floor-facing · invisible to canine vision · low-light camera assistSK6812 RGBW LED RING12 LEDs · NeoPixel single-wire · dims OFF during dog sessions · 74AHCT125 level-shiftUSB-C 5V/3A MAINS INLETNo battery — mains-only for session-timing reliability · silicone gasketSAGE TPE TREAT TRAY (EXTERNAL)Food-grade Shore 40A · 64 mm ext. · removable · dog approaches to eat — approach IS the protocolLegendStructure / chassisMotor / actuatorCamera moduleSensorMain PCBSpeaker driverSilicone / trayPawvlov · Rev A · template

S/L variants: same PCB and electronics; enclosure height and hopper volume scale. Tray is external to the sealed chassis — the dog must walk toward the device to retrieve the treat, mechanically reinforcing approach behavior as part of the DACVB counter-conditioning protocol.

Chip selection

ESP32-S3 chosen for integrated WiFi + BLE + dual-core 240 MHz with SIMD vector instructions for on-device INT8 inference at ~$3.85 per chip @1K MOQ ($3.20 @10K).

Sensor & camera architecture

Floor-level 120-degree wide camera reads dog body language; 4-mic PDM MEMS array detects low-frequency thunder and classifies storm vs household noise; IMU detects tip-over for safety lockout.

  • OmniVision OV5640 5 MP CMOS camera + 120-degree wide lens
  • InvenSense ICS-40180 PDM MEMS microphones (4 array)
  • Bosch BMI270 6-axis IMU
  • 940 nm IR-assist LED ring (8 LEDs, floor-facing)

ML inference pipeline

Two-stage inference: on-device MobileNet v3-Small classifies dog state every 500 ms; cloud Gemini Vision fallback handles borderline cases (confidence < 0.7) when WiFi is connected.

Power & battery

USB-C mains powered for session reliability; no battery in the dispense path because a brownout mid-session would mis-time the treat and risk fear reinforcement.

Firmware architecture

FreeRTOS dual-core split: mic + inference + state machine on core 1, WiFi + BLE + cloud sync on core 0. Hardware watchdog + dispense safety lockout on motor stall current.

Manufacturing process

  • Injection mold (4-cavity ABS shell, separate molds for S / M / L)
  • Overmolded TPE silicone treat tray (food-grade, FDA-compliant)
  • SMT PCBA line + AOI inspection on the 4-layer 1.8 mm PCB
  • Functional end-of-line test (mic frequency sweep, camera focus check, auger cycle, BLE/WiFi RF validation)
  • Acoustic chamber QC sample audit confirming < 40 dB at 1 m on dispense

Architecture summary

ESP32-S3 dual-core MCU drives a 4-MEMS PDM microphone array (storm detection + classification at 80 Hz) plus an OV5640 120-degree wide camera running on-device MobileNet v3 INT8 inference for dog-pose classification (calm | alert | anxious | absent), plus a near-silent N20 12V gear motor + Delrin helical auger for treat delivery (<40 dB at 1 m). The companion mobile app pairs over BLE and syncs telemetry over WiFi; a thin cloud control plane (FastAPI + Postgres on Render, R2 object storage for 30-sec session clips) handles cloud-fallback inference, OTA firmware, Stripe billing, and session-history aggregation. Mains-powered via USB-C - no battery in the dispense path so brownouts cannot mis-time a treat during a session.

Risk register

Risk register

RiskSeverityLikelihoodMitigationConfidence
On-device dog-pose classifier accuracy drops on dark-coat breeds in low-light home environments; a calm-state misclassification = treat during fear state = iatrogenic outcome and brand-collapsing viral failure threadHighMedium940 nm IR-assist LED ring (invisible to dogs) extends low-light range; confidence threshold 0.7 + cloud-fallback inference for borderline cases; Phase 2 beta data collection deliberately oversamples dark-coat breeds and dim rooms; quarterly OTA model updates Modeled
OV5640 camera EOL or supply disruption mid-production; sole-source for the primary inference sensorMediumMediumQualify pin-compatible second source (ARC0143 / SC130GS) before the 10K production run; maintain 6-month buffer stock at 1K MOQ Sourced
FCC Part 15C intentional-radiator certification slips past launch window due to BLE+WiFi co-existence RF emissionsMediumLowPre-scan at design freeze with in-house spectrum analyzer; engage Sporton lab early for Phase 3 Month 1 submission; $3K remediation reserve inside capital_ask Modeled
TPE-to-ABS overmold adhesion fails after 6+ months of treat-residue cleaning cycles; treat tray separates from chassis under field useMediumMediumMaterial supplier DFM review at PCB rev 2; accelerated wear test (1,000 cleaning cycles) before committing production tooling; 12-month replacement warranty reserve in COGS Assumption
Auger motor stalls on oversized treats jamming the helical screw; a mid-dispense stall = unfinished session = wrong association reinforcedMediumMediumStall-current detection triggers safe-stop + reverse-jog jam-clear routine; user-facing 'recommended treat size' guidance on packaging + app; physical hopper-feed geometry rejects oversized treats at the throat Modeled
Injection-mold tooling lead time exceeds 12 weeks (Shenzhen partner), pushing launch to Month 24+HighMediumLock partner + lead-time commitment in writing by Phase 2 Month 2; stage 30% deposit on Phase 3 go-signal; arrange express freight reserve ($5K) if needed to compress timeline Sourced

Dimensions (to scale)

Overall dimensions by model

ModelHeight (mm)Width (mm)Depth (mm)Weight (g)Note
Small142127127480Breeds under 20 lbs; 60-treat hopper; 120g steel ballast; 50mm tray
Medium165140140620Breeds 20-60 lbs; 180-treat hopper; 200g ballast; 64mm tray — director baseline
Large200161161880Breeds over 60 lbs; 400-treat hopper; 280g ballast; 81mm tray; taller hopper chamber

Component dimensions (to scale)

ComponentValue (mm)ToleranceNote
Bayonet lid ODS:89 / M:102 / L:117±0.3mmQuarter-turn child-proof push+turn; flush with top face
Hopper chamber heightS:52 / M:70 / L:95±0.5mmCapacity 60 / 180 / 400 treats by size
Camera bore OD15±0.1mmOV5640 module seat + brushed aluminum ring; shared all sizes
Camera lens FOV120 degrees (all sizes)±2 degreesFixed wide lens, no mechanical zoom
MEMS mic spacing12±0.2mmCenter-to-center x4 mics, horizontal array, flush front face
Acoustic foam chamber wall12±0.5mmClosed-cell melamine foam; 35-40 dB dispense attenuation
Auger helical pitchS:9 / M:14 / L:19±0.2mmOne full helix per treat cell; per-size Delrin POM part
Auger throat diameterS:8 / M:13 / L:18±0.1mmMax treat OD = throat minus 1mm clearance
Funnel angle35 degrees from vertical (all sizes)±1 degreeGravity feed; prevents treat bridging at the throat
Silicone tray extensionS:50 / M:64 / L:81±0.5mmForward from chassis front face; per-size TPE part
Silicone tray well depthS:8 / M:10 / L:12±0.3mmPrevents treat roll-off; Shore 40A TPE
Speaker aperture OD38±0.3mmBehind front lower mesh grille; shared all sizes
LED ring orbit diameterS:100 / M:118 / L:135±0.5mmSK6812 RGBW under-glow ring at base perimeter
USB-C port recess depth3±0.2mmRecessed with silicone gasket; shared all sizes
ABS wall thickness2.5±0.1mmStructural minimum for 1m drop test; all sizes
PCB dimensions100 x 70±0.1mm4-layer 1.6mm FR4; shared across all three size variants
Base ballast disc ODS:70 / M:85 / L:100±0.5mmSteel disc; S=120g / M=200g / L=280g embedded in ABS base plate

Inside the device

Cutaway view with labeled internal components