SentinelBox

Predictive maintenance at the edge for HVAC systems — 🥇 1st place, Keep It Cool Hackathon (Carrier, 2026).

Every HVAC already measures current and temperature to operate. Nobody uses that data to prevent failures. SentinelBox taps into the signals every HVAC already has, complements them with new sensors integrated in the box (vibration), and learns each unit's normal behavior — so maintenance happens before performance drops, not after.

The problem

Most HVAC failures are detected when performance has already degraded. There is a window of opportunity — the first early signals — that goes unnoticed today. And a single fixed threshold for every unit makes it worse: false alarms trigger unnecessary visits, missed detections end in breakdowns.

The solution

SentinelBox learns each unit's normal instead of applying one threshold to all of them. Two demo fans intentionally run different normal current profiles; the system still recognizes each one's normal independently.

• Edge processing on ESP32, no cloud required. The decision runs on the gateway.
• One baseline per unit (current, temperature, vibration), learned in ~90 s.
• Two states: ● NORMAL ● MAINTENANCE REQUIRED (+ WARNING while degrading, LEARNING while baselining).

Edge AI: a neural network running on the ESP32

The gateway doesn't apply fixed thresholds — it runs a neural network, on-device, every 2 seconds. No cloud, no round-trips, no connectivity required.

1. Sense. Each 2 s window: 100 vibration samples @ 50 Hz from the MPU6500 (±4g, per-axis means removed) → RMS, crest factor, kurtosis. Mean current (INA219) and temperature (MAX6675) arrive from the HVAC node over UART.

2. Compare against a learned baseline, not programmed constants. The first 45 windows (~90 s) of normal operation become that unit's baseline, persisted in NVS flash so it survives reboots (r over serial relearns it). Every network input is expressed relative to baseline — so one trained network works across fan sizes, speeds, and mountings:

#	Network input	Formula
1	vib_rms	log₂(RMS / baseline)
2	vib_crest	log₂(crest / baseline), faded out near the noise floor
3	vib_kurt	log₂(kurtosis / baseline), faded out near the noise floor
4	current	log₂(I / baseline)
5	temp	(T − baseline) / 5 °C

3. Infer on-chip. MLP 5→16→8→3 (ReLU, softmax T=3.0): 259 float32 parameters ≈ 1 KB, written in plain C (SentinelBox/sentinel_ann.h, weights in generated SentinelBox/sentinel_model.h). No TensorFlow Lite, no heap allocator — deterministic inference in microseconds on the ESP32. Outputs NORMAL / WARNING / MAINTENANCE, converted to health % rescaled so the unit's own baseline reads exactly 100.

4. Explain + persist. An attribution pass re-runs the net with the other sensors held at baseline, so the dashboard can say why ("Cause: vibration, +0.18 mm/s vs normal (+16%)"). Health is EMA-smoothed (α=0.35) and the status flips only when the abnormal class persists 3 of the last 5 readings — one isolated peak changes nothing.

Measured model performance (ml/artifacts/metrics.json, held-out test set):

	Detection rate	False alarm rate	Missed anomaly rate
Neural network	1.000	0.003	0.000
Weighted-score rule (baseline)	0.965	0.0003	0.035

Overall test accuracy 0.9936; C/Python inference parity verified with gcc (max probability diff ~1e-6). Retrainable on your own recordings — see docs/setup.md and ml/README.md.

Why edge instead of cloud: works with zero connectivity (the demo runs on a phone hotspot with no data plan); milliseconds from sense to decision; no bandwidth or cloud bills per unit; raw vibration never leaves the site; behavior is deterministic and testable on the bench.

Validation results

Measured on the prototype (normal operation vs. controlled perturbations: hair dryer for temperature, airflow/rotor interference for current and vibration):

Metric	Result
Average detection time	6 s
Baseline learning time	90 s
False alarm rate	0 %
Detection performance — current / temperature / vibration	6 / 3 / 57

System architecture

┌──────────┐  UART JSON   ┌─────────────┐  WiFi (iPhone hotspot)   ┌──────────────┐
│ HVAC-01  │ ────────────▶│             │ ────────────────────────▶│              │
│ INA219 + │  temp+curr   │ SentinelBox │  POST /api/readings      │  Next.js app │
│ MAX6675  │              │ ESP32+MPU   │  (+ vibration measured   │  SQLite +    │
├──────────┤  (plug &     │ gateway     │   on the gateway)        │  3D dashboard│
│ HVAC-02  │   play, one  │             │                          │              │
│ INA219 + │   at a time) │             │                          │              │
│ MAX6675  │              │             │                          │              │
└──────────┘              └─────────────┘                          └──────────────┘

• HVAC nodes (HVAC01/, HVAC02/) — sensor boards on each fan. They only send what they measure (unit_id, temperature, current) plus a HELLO with their ID on boot. No WiFi, no decisions.
• SentinelBox gateway (SentinelBox/) — one UART port, WiFi, and its own vibration sensor. It completes the 6-field reading, runs the neural network, and POSTs to the server. Swap HVACs live and it detects the new unit by itself (retrofit plug-and-play demo).
• Web app (src/) — Next.js dashboard: 3D building twin with clickable units, live charts, alert cause breakdown, and a REST API with SQLite storage. Works fully offline on a local network.

Business model

A service inside the policy the customer already pays. Installed by Carrier technicians during an already-scheduled visit — no extra trip.

Upfront (device + install, once)	$3,000 MXN
Monthly (dashboard, alerts, support)	$150 MXN
Sold to	Stores with rooftops, Carrier BluEdge policy customers in Mexico
Margin for Carrier	$1,267 per install / $121 monthly per unit

Projected 5-year return: 3.9x ROI · 38% IRR · $2.76M NPV. Next step: finished product and pilot with Carrier policy customers.

Try the demo (60 seconds, no hardware)

npm install
npm run dev          # dashboard at http://localhost:3000
npm run sim:seed     # backfill 25 min of both units
npm run sim:start    # stream live readings every 2 s

Open the dashboard: HVAC-02 degrades live while HVAC-01 stays healthy — each against its own baseline.

Full technical setup

All engineering docs live in docs/:

• docs/setup.md — hardware, pin tables, firmware flashing, network, dashboard
• docs/api.md — REST API reference with curl examples
• docs/demo-checklist.md — field day-of-demo checklist (hotspot, IPs, flash order, debugging)

Repository layout

Path	What
SentinelBox/	Gateway firmware (ESP32 + MPU6500 + embedded ANN)
HVAC01/	HVAC-01 node firmware (ESP32-C3, PWM fan, INA219 + MAX6675, UART)
HVAC02/	HVAC-02 node firmware (ESP32, fixed fan, INA219 + MAX6675, UART)
src/	Next.js dashboard + REST API (src/app/api/*)
src/lib/sentinel/	Shared validation, SQLite store, simulator, API client
ml/	Neural-network training pipeline (retrainable, exports to ESP32)
data/	Local SQLite database (git-ignored, auto-created)
docs/	Technical setup, API reference, demo checklist

Team

Eduardo Pérez · Pedro Uribe · Alejandro Chio · Emiliano Méndez · Miguelangel Rodríguez — 🥇 1st place, Keep It Cool Hackathon (Carrier, 2026).

Repository: https://github.com/ayexxxd/SentinelBox