A crash on an empty road shouldn't wait for a passer-by.
MRAI is a small node on a lamp post. Cheap sensors watch the lane all the time; when something goes wrong, the camera and AI wake up, check, and alert a human operator's phone — in about a second on the bench.
- ≈1.2 s
- Sensor → alert
- Measured · v1 bench
- ₹8k
- Main node / year
- Indicative
- 0
- Video uploaded
- By design
- 2×
- Runner-up finishes
- Competitions
Pings the lane 20 times a second. A sudden drop wakes the camera.
From a distance drop to a ringing phone.
The AI sleeps until it's needed.
ESP32 reads distance, rain and smoke 20 times a second on milliwatts.
A sudden distance drop or smoke spike raises a wake signal.
The Pi wakes the camera and grabs 3–5 frames.
YOLOv8n checks the frames at 2–4 fps and confirms or dismisses.
Why it matters: a bare Pi 3B can't run continuous video AI. Gating means it doesn't have to — and the node uses a fraction of the power. Timings marked indicative are design targets, not field measurements.
What we built, and what a city would buy.
| Part | Bench unit · v2 (built) | Road-ready unit (target) |
|---|---|---|
| Compute | Raspberry Pi 3B + ESP32 | Pi 4/5 + Coral or Hailo-8L AI chip |
| Video AI | 2–4 fps, triggered bursts | 20+ fps continuous |
| Distance | HC-SR04 (indoor) | JSN-SR04T waterproof |
| Enclosure | Open breadboard | IP66 sealed box |
| Power | Mains adapter | 12 V LiFePO4 + solar panel |
| Radio | LoRa SX1276 | LoRa + 4G on the main node |
| Parts cost | ≈ ₹8,000 (indicative) | ≈ ₹28,000 (indicative) |
Built for cities and for companies.
- Start with a supervised corridor pilot — Wazirabad Marg, 10 nodes.
- Alerts in a standard format (CAP) for city control rooms and 112.
- Human verifies every alert before anyone is dispatched.
- No video stored — fits India's DPDP Act 2023 by design.
- One-time price per node, with support and alert routing.
- Incident records and CSV exports for safety audits.
- Simple API to plug alerts into existing systems.
- Supervised private-site trial before any contract.
Where MRAI actually stands.
Live. Node registry, incident triage, CSV exports, map and phone alerts all work.
v2 built and bench-tested, then taken apart. Paused for time and funding.
Funding for one road-ready node and a consenting trial site to collect real evidence.
What people ask us.
Is there a working device right now?
Prototype v2 was built and bench-tested, then taken apart for lack of time and funding. The software platform is live; what you see here runs on simulated readings until a node is rebuilt.
A Raspberry Pi 3B only does 2–4 frames a second. Isn't that useless?
Not with sensor gating. The cheap sensors watch the lane all the time; the camera and AI only wake when the distance sensor sees something wrong, and then check a short burst of frames. That's how roadside units save power. A road-ready unit adds an AI chip (Coral or Hailo) for continuous video.
What happens if the internet goes down?
Detection is on the node. Alerts hop over LoRa radio to the main node, and packets are stored and sent again once the link returns.
Does it call the ambulance itself?
No. MRAI informs a human operator, who checks and then calls 112. Automatic dispatch is out of scope until an authorised trial proves it.
What about privacy?
No video leaves the node. Only the event — type, time, place, sensor readings — is sent. No faces, no number plates.
Who is behind it?
UniWebs, a student-run engineering studio founded by Kavy (software) and Saksham (hardware), both in Class 11. MRAI finished first runner-up twice in competition.