☁
Bad air
Diesel exhaust (CO, NO₂), blasting gases, dust, and falling oxygen. Often you can't smell the danger until it's serious.
Cursor Hackathon Sudbury · Track 03
Build the North · 2026
The offline AI oracle for underground mine intelligence.
OreAcle watches a mine's sensors, detects danger before a human would, explains it in plain English, recommends the action, and drives the ventilation fan itself, all running locally on a laptop at $0 AI cost.
$0
AI running cost
100%
Open-source & local
2
AI techniques (ML + LLM)
$40
Edge node (Pico W)
The idea, in plain language
The mine already measures everything. It just isn't understood fast enough.
Two kilometres underground, three things quietly cause harm: bad air, failing equipment, and mismatched ventilation. None announce themselves clearly. Today a human is expected to watch dozens of live numbers across many zones and connect the dots, every time, all shift. That's exactly the job a machine should help with.
⚙
Failing equipment
Pumps, fans, and drills vibrate differently for days before they fail. Miss it, and a $500K breakdown, or worse, follows.
🌀
Wasted ventilation
Fans can be 40%+ of a mine's energy bill. Too much wastes money; too little is a safety risk. Matching airflow to people is hard by hand.
The gap we close
The sensors exist. The brain and the friendly face don't.
Maestro Digital Mine has air-quality stations running in Vale's Sudbury mines (including Creighton, 2.4 km deep) since 2015. Sandvik's Newtrax platform monitors Glencore's people and equipment underground. Vale runs a remote operations centre in Sudbury watching all five Ontario mines, with ventilation-on-demand saving millions.
So the data is there. What's missing is the AI interpretation layer that turns raw numbers into a prioritised, plain-English decision, and a practical operator interface built for the person on shift, not just an engineer. OreAcle is that missing layer, built with free open-source AI.
"Vale and Glencore have the eyes and the muscles. OreAcle is the open-source, zero-cost brain and friendly face that connects them, proved end-to-end on a $40 board."
Sense
→
Think
→
Explain
→
Act
The whole story is one round trip: feel the world, decide, explain it to a human, and turn the fan, closing the loop.
System clarity
Who uses OreAcle, and how does it actually work?
Two kinds of people use OreAcle at different locations. Here's exactly what each person sees on their screen, and where every piece of hardware, software, and AI runs.
Underground Operator
On-site worker, up to 2 km underground
Location: active stope, equipment bay, or shaft entrance
Carries a tablet or ruggedized iPad on the job. Their key question: is my zone safe right now, and what should I do?
- Zone status: OK, WATCH, or ALERT — colour-coded and instant
- Plain-English AI alert: "CO at 45 ppm and rising — increase ventilation now"
- Live readings: CO, NO₂, temp, humidity, vibration, airflow
- AI Copilot: ask "Is it safe to bring more workers in?" and get a direct answer
- Fan speed readout and manual override control
Surface Supervisor / Shift Manager
Remote monitor, surface control room or office
Location: surface operations centre or any networked laptop
Sits at a desktop or laptop. Needs to watch all zones at once, catch problems before they escalate, run training scenarios, and control ventilation across the mine.
- All 4 mine zones on one screen with live status dots
- 60-second trend charts for gas levels, vibration, and temperature
- AI alert feed: what triggered, when, and the recommended response
- Manual fan override and 5 demo scenarios for training and testing
- Full AI Copilot chat for post-incident analysis and shift briefings
How data flows, end to end — 5 steps
1
Sense
Pico W edge node reads sensors ~2× per second
2
Stream
JSON over USB serial from Pico W to MacBook backend
3
Think
MacBook runs the AI brain — fully offline, $0 cost
4
Show
WebSocket pushes live data to any browser in real time
5
Act
AI commands the fan via WebSocket back to the Pico W
What's real hardware vs simulated (SIM):
DHT11 (temperature + humidity) and MPU6050 (vibration/tilt) are real sensors being assembled on the Pico W.
CO, NO₂, PIR occupancy, and airflow are SIM software values in this POC —
swapping in a ~$10 MQ-7 gas sensor is a single wire change.
The AI brain (FastAPI + Isolation Forest + Ollama), the WebSocket, and the browser dashboard are 100% real and running live right now.
How it works
System architecture
The Pico W is the nervous system: it feels the world and reports back over a USB cable. Your MacBook is the brain: it thinks, decides, explains, and shows everything on screen, then sends a command back to spin the fan. Everything below runs locally and free.
Underground
Equipment + Haul Truck
Miner
Gas / Air Hazard
OreAcle Edge Node
Surface Control Room
OreAcle Offline AI
Offline
$0 AI
Open-source
Shift Supervisor
Automated Response
Ventilation Fan
Local Alarm + LCD
Technical detail
System architecture
🔌
Edge Node
Raspberry Pi Pico W · MicroPython
MPU6050vibration / tilt
DHT11temperature · humidity
PIR + Ultrasonicworker occupancy
Buzzer · RGB · LCDlocal alarm & readout
DC motor "fan"driven by AI command
→
JSON
USB serial
USB serial
💻
The Brain: MacBook
all local · $0 AI cost
Ingest → SQLitetime-series store
Isolation ForestML anomaly detection · scikit-learn
Rules + Ventilation-on-Demandsafety limits · fan setpoint
Ollama LLMLlama 3.2 3B · alerts + copilot
FastAPI + WebSocketlive push to UI
⇄
WebSocket
🖥️
Operator UI
Next.js · React
Live gauges + zone mapglanceable status
AI alert feedplain-English warnings
AI copilot chatask in natural language
Ventilation controlfan spins on screen and in real life
Sense → Think → Explain → Act
•100% local
•open-source
•$0 AI cost
•works offline (no connectivity needed underground)
Data contract. The edge node emits one JSON object per reading (~2/sec) that flows unchanged through ingest → DB → ML → LLM → UI:
{ "ts": 1750550400, "zone": "Z3", "vib_rms": 0.07, "temp_c": 24.1,
"humidity": 58, "occupancy": 2, "distance_cm": 180,
"co_ppm": 12, "no2_ppm": 0.4, "airflow": 0.42, "fan_pct": 40 }co_ppm, no2_ppm, and airflow are software-simulated in the POC; swapping in a ~$10 gas sensor is a one-wire change.
What we use, and how little
Hardware: humble, plug-and-play, in the box
Everything comes from one SunFounder Raspberry Pi Pico W Ultimate Starter Kit. No soldering. Hardware is ~10% of the effort; the idea, the AI, and the interface are the stars.
RP2040
Raspberry Pi Pico W
The edge "nervous system"
- CPUDual-core Arm Cortex-M0+ @ 133 MHz
- Memory264 KB SRAM · 2 MB flash
- Wireless2.4 GHz Wi-Fi (CYW43439)
- LanguageMicroPython
- Link to MacUSB serial (primary)
Myth-buster: Amazon lists "Tegra / FreeRTOS", that's an auto-fill error. It's an RP2040 microcontroller. It's too small to run the AI, which is exactly why the brain lives on the MacBook.
| Role in OreAcle | Component | Interface |
|---|---|---|
| Equipment vibration / ground movement | MPU6050 | I²C 0x68 |
| Air temperature + humidity (heat-stress) | DHT11 | 1-wire |
| Worker occupancy in zone | PIR motion | digital |
| Distance / presence backup | Ultrasonic HC-SR04 | trig/echo |
| Local danger alarm (sound) | Active buzzer | PWM |
| Status light (green/amber/red) | RGB LED | PWM |
| Local readout ("VENT ↑") | I²C LCD1602 | I²C 0x27 |
| The "ventilation fan" (the wow) | DC motor + driver | PWM |
Extra spend on AI & hardware to make this run: $0: it all runs on a laptop and gear mines already own.
The AI: open-source, local, free
Two kinds of AI, zero dollars, no internet required
Everything runs on the laptop with no API keys and nothing leaving the machine, which matters underground, where connectivity is poor and data is sensitive.
Detection
Isolation Forest
scikit-learn
A lightweight, explainable machine-learning model that learns what "normal" looks like and flags abnormal combinations of readings, even when no single number crossed a hard limit.
- Instant on CPU
- Explainable
- Backed by hard safety rules
Explanation + Copilot
Llama 3.2 (3B)
via Ollama · runs locally
A small open-source large language model turns each anomaly into a one-sentence alert with a recommended action, and powers a chat copilot you can ask in plain English.
- $0 / offline
- Tuned for an Intel-i9 / 16 GB Mac
gemma2:2bfast fallback
Optional
faster-whisper
speech-to-text
A stretch goal that lets a supervisor speak a question hands-free ("is Zone 3 safe?"). Runs on CPU, fully local.
- Hands-free
- Local
- Add only after core is done
$0
No API bills, ever. The model files are downloaded once and run forever for free.
⛓️💥
No internet needed at demo time. Critical underground and for data privacy.
⚡
Small models keep alerts snappy; outputs are kept to ≤ 2 sentences and pre-warmed.
Proven in the field
Industry in action
OreAcle is the open-source AI interpretation and operator-interface layer that complements these proven deployments. The sensors, the data, and the infrastructure already exist underground.
Air Quality Monitoring
Maestro Digital Mine
Vigilante AQS in Vale's Sudbury Mines
Hundreds of Vigilante AQS air-quality stations have been running in Vale's underground operations since 2015, covering airflow, gas, temperature, and dust. Deployed at Creighton Mine (2,400 m deep), one of the deepest in Canada.
Fleet Telemetry & Safety
Sandvik + Glencore
Newtrax Collision-Avoidance Expansion (July 2025)
Sandvik and Glencore expanded their partnership in July 2025 to deploy Newtrax collision-avoidance and the Mining Data Platform across Glencore's mixed underground fleet at Raglan (Canada) and Kamoto (DRC).
Remote Operations
Vale IROC + VOD
Sudbury, Ontario
Vale's Integrated Remote Operations Centre (IROC) in Sudbury oversees all five Ontario underground mines from surface. Ventilation-on-Demand systems save millions of dollars annually by matching airflow to activity.
Air Quality Monitoring
Accutron Instruments
MAQS Mine Air Quality Station
The MAQS station monitors gas concentrations, airflow, and environmental conditions in Northern Ontario mines. A well-established Sudbury-region instrument measuring the same parameters OreAcle interprets.
Hackathon Sponsor
Key Logic Software
Sudbury, Ontario: software wrapped around mining hardware
Key Logic builds embedded systems, AR/VR (VirtualMine), and custom sensor-data algorithms for mining clients including Minewise ShafTrack, GeoSight NX-150, and TesMan remote loading.
OreAcle's role: These companies excel at collecting data and keeping workers safe physically. OreAcle is the open-source AI layer that adds plain-English interpretation and a human-centred operator interface on top, vendor-neutral across all their sensor networks.
See it before the hardware arrives.
The Live Console below runs the entire experience on sample data. Trigger a gas build-up or a bearing failure and watch the AI alert, the fan response, and the copilot react in real time.
Zones
live statusVentilation control
demand-based40%
Sensors: Entry / Shaft
real-timeTrends
last 60sAI alert feed
LLM-writtenAI Copilot
ask in plain EnglishBackend offline — copilot is answering from the built-in simulation.
Rendering the master project plan…
Cursor Hackathon Sudbury · Track 03 · Build the North 2026
Build Tracker
A transparent, judge-friendly view of implementation progress. All features listed here are working in the live demo right now unless otherwise noted.
Backend: Running
WebSocket: Live
Frontend: Connected
Hardware POC: Active
Current Build Status
Demo-ReadySoftware MVP complete · Hardware POC in progress
All four mine zones stream live data end-to-end. Anomaly detection, VOD fan logic, LLM alerts, and copilot are fully operational. Pico W sensor node hardware POC is now being assembled.
Completed Milestones
Backend & API
- DoneFastAPI app with
GET /health,WS /ws,POST /scenario/{name},POST /fan,POST /copilot - DoneWebSocket streams all 4 zones at 0.5 s cadence with 14 sensor fields per reading:
ts, zone, vib_rms, temp_c, humidity, occupancy, distance_cm, co_ppm, no2_ppm, airflow, fan_pct, status, reasons, alert_text - DoneSQLite readings log at
data/oreacle.dbwith index on zone and timestamp - DoneMakefile:
make install,make backend,make sim,make web
Simulation & AI Layer
- DoneFour-zone mine simulator with per-zone realistic baselines, Gaussian noise, and occupancy variation
- DoneFive scenarios:
normal,gas_buildup,bearing_failure,blast_clearance,reset - DoneIsolation Forest anomaly detection (scikit-learn) trained on 2,000 synthetic samples across all 4 zones
- Done10-rule safety threshold system covering CO, NO₂, vibration, temperature, and airflow
- DoneVentilation-on-Demand (VOD) fan setpoint: base + gas demand + occupancy + airflow compensation + temp + status floor
- DoneLLM alert generation via Ollama (
llama3.2:3b) with deterministic template fallback when Ollama is offline
Frontend Integration
- DoneWebSocket connection with 3 s auto-retry; "Connected / Backend offline" status badge with pulsing dot
- DoneLive zone cards for all 4 zones showing status, CO, NO₂, vibration, temp, humidity, airflow, occupancy, fan %
- DoneSix live gauges for active zone, 60-second trend chart, and per-zone AI alert feed with deduplication
- DoneAI Copilot chat — POSTs to
/copilotwhen backend is live; template fallback when offline - DoneFan slider with Auto (VOD) / Manual override modes; manual mode POSTs to
/fan - DoneGraceful offline fallback: built-in JS simulator mirrors backend baselines so the console never crashes
In Progress
Active- In ProgressHardware POC Phase 1: Pico W + DHT11 + MPU6050 sensor node streaming USB JSON to backend
- In ProgressDemo rehearsal: end-to-end scenario walkthrough for judges
- In ProgressSite polish and pitch deck preparation
Hardware POC
In ProgressThe SunFounder Raspberry Pi Pico W Starter Kit is being used to build a working edge sensor node. The goal is to replace the software simulator with real sensor readings streamed over USB JSON to the MacBook backend. All backend ingest code is already written and ready to receive live serial data from the Pico.
Phase 1 — In Progress
- In ProgressRaspberry Pi Pico W edge node flashed with MicroPython firmware
- In ProgressDHT11 temperature and humidity sensor wired and reading
- In ProgressMPU6050 vibration and tilt sensor wired via I²C (address 0x68)
- In ProgressUSB JSON stream from Pico W to MacBook backend via
pyserial - In ProgressRGB LED status indicator driven by zone safety status (green / amber / red)
Phase 2 — If Time Permits
- PendingDC motor fan controlled by AI-commanded fan % setpoint (physical demo wow-factor)
- PendingI²C LCD1602 local status readout ("VENT UP", "CO ALERT")
- PendingActive buzzer local alarm triggered on ALERT status
- PendingPIR motion sensor + HC-SR04 ultrasonic distance for occupancy detection
Demo Readiness
Ready NowEverything below works today on a laptop with no hardware and no internet connection. Run make backend in one terminal and open the Live Console.
1
Start the backend:
make backend — FastAPI starts on :8000, simulator begins streaming immediately.2
Open the Live Console. The status badge flips to "Connected" and all 4 zone cards populate with live data within one second.
3
Press "Gas Buildup". CO rises in Zone B. The Isolation Forest flags the anomaly. The VOD logic ramps the fan. An LLM-written alert appears in the feed.
4
Press "Bearing Failure". Zone C vibration spikes toward the ALERT threshold. Status switches. A new alert describes the mechanical risk and recommends action.
5
Ask the AI Copilot: "Which zones are unsafe and why?" The LLM answers using the last 20 real readings from SQLite.
6
Press "Reset". All zones return to normal baselines. The alert feed clears. The fan drops to the base VOD setpoint.
Demo Recording
Watch Console Demo (MOV)
Recorded 2026-06-27 · Live console + backend walkthrough · local playback
Next Steps
In Progress
Hardware POC
Pico W potentiometer edge node is live. Turn the knob to drive a calibrated airflow signal into the dashboard in real time.
⚡ Open Hardware POC →Stretch
Voice Copilot
Add faster-whisper for speech-to-text so judges can ask the AI copilot hands-free: "Is Zone B safe right now?" Runs locally, no API key needed.
Stretch
Per-zone Fan Control
Individual VOD fan setpoints per zone. The API already accepts a zone parameter on POST /fan; the UI slider only sets globally today.
Stretch
Data Export
Download the full readings log as CSV from the Live Console. SQLite at data/oreacle.db is already logging every reading; the export endpoint just needs to be wired up.
Hardware Proof-of-Concept · Raspberry Pi Pico W
Live Hardware POC
Raspberry Pi Pico W edge node streaming calibrated airflow telemetry over USB. Plug in the Pico, click Connect, and turn the knob.
Not connected
Calibrated Airflow Signal
0.00
m/s
Live Pico W potentiometer signal, smoothed for demo stability.
--
Ventilation Fan
--%
Fan % increases as airflow decreases — mirrors VOD logic.
Sensor Details
About This Demo
The physical potentiometer is a calibrated stand-in signal for a real airflow sensor. It is not measuring actual airflow — it demonstrates the full data pipeline: ADC reading → USB JSON → Web Serial → live dashboard update.
Signal processing: 50-sample ADC average → exponential smoothing (α = 0.08) → deadband filter (±0.10 m/s) → clamped 0–5 m/s range. Status thresholds: ≥ 1.7 m/s → OK · ≥ 0.8 m/s → WATCH · below → ALERT. Fan % rises as airflow falls.
Firmware: hardware/firmware/main_potentiometer.py — upload to Pico as main.py via Thonny.
Requires Chrome or Edge over http://localhost. Close Thonny first (port conflict), then click Connect to Pico and pick the usbmodem port.