[04]hackathon

★ First all-nighter

AI for Sustainability Hackathon 2026

Participant

The hackathon where we went properly deep into research. We built Fractal Fire Mamba, a forest fire detection network that uses state space models and chaos theory to tell a real fire from fog, dust or car exhaust, and keeps alerting even after the fire destroys nodes. We did it on our first all-nighter, half asleep in the university library.

When
February 2026
Where
Online · Canadian University Dubai, UAE
Hosted by
Canadian University Dubai · American Society for Engineers
Format
online · 3 days
Track
Forest fire detection & prevention
Evolve AI at AI for Sustainability Hackathon 2026

[01]The squad

Evolve AI

  • Govind Jindal← that's me
  • Piyush Sharma
  • Bhavishya Grover
  • Karunika Chaudhary
  • Anayat Virk
  • Deepanshu Arora

My role · Backend, cloud & strategy

[02]The rounds

  1. Round 1

    The deck

    An online idea round: our problem statement, the 6-phase detection pipeline, the hardware and its costs, all in one deck. We submitted it twice (see the side quests).

    Selected ✓ ×2

  2. Round 2

    The pitch

    An online pitch to the jury. Take one, on no sleep, did not go well. Take two, with our second team number, did.

    Take 2 ☕

  3. Round 3

    The build

    Three days online, 14–16 February: the full pipeline, the multi-queen mesh and the Fractal Sentinel command dashboard, live on Leaflet maps.

    📜 Participated

[03]What we built

Fractal Fire Mamba

Fractal Fire Mamba home page

  • Sensor fusion + Mamba SSM

    Chemical (50%), visual (30%) and environmental (20%) signals fused by a state space model with a 60-second memory, so it reads the trend and not a single snapshot. Linear time, not quadratic attention.

  • Watchdog layer

    Rejects impossible readings, frozen sensors and drift, and fills gaps with virtual sensors, so "ghost alarms" never reach the model.

  • Fractal gate

    Rescaled-range analysis gives a Hurst exponent. H > 0.6 means structured, persistent signal (a fire signature) instead of random noise, and only then does the camera wake up.

  • Chaos kernel

    A Lyapunov exponent from phase-space embedding. λ > 0 means self-sustaining, exponential growth, the difference between a spreading fire and a passing diesel truck.

  • Wake-on-event vision

    RGB by day and an MLX90640 thermal camera at night (23 mA against 250 mA), powered only when chaos is confirmed. That saves about 800 mAh a day per node.

  • Logic gate + witness protocol

    Four alert tiers from green to red. Before red, neighbouring nodes within 500 m must confirm it. No single node can start a panic.

  • The Hive

    $160 drone nodes and $700 queen nodes on a self-healing LoRa mesh, with Iridium satellite uplink for critical alerts and death-vector maps of where the fire will spread.

Mamba SSMPyTorchFastAPINumPy & SciPyOpenCVLeaflet.jsChart.jsLoRa meshIridium satelliteRaspberry Pi

GitHubProject videoLinkedIn postEvent

[04]The film 🎬

[05]Side quests

  • Registered twice. Selected twice.

    We accidentally registered twice and ended up with two team numbers, #5208 and T144. We sent the round-one deck in under both… and both got selected.

    Selected ×2

  • The pitch that wasn't

    After a night with no sleep, our first online pitch did not go well. Like, at all.

    Not our finest

  • Take two

    Plot armour: a spare team number. So we used it, and pitched the whole thing again, properly this time.

    Redemption ✓

  • An AI-made film

    We made the project video with AI, somewhere between commits. Watch it above. 🍿

    AI cut

[06]The night shift 🌙

Commits
13
Span
13h 41m
Avg. message
55.8 chars
Shortest
“theme”

Our first all-nighter, in the university library, half asleep. The git log keeps the receipts: a merged pull request at 1:31 AM, a new model at 3:11, a "theme" at 4:13, and fire-spread vectors by breakfast. 🔥 Four months after AyurSutra's "c", the commit messages finally had words in them. 📈

[07]From the floor

4:30 AM. Between "theme" (4:13) and the next push (5:29). 🫠
Risk trend, Hurst and Lyapunov, live
Risk trend, Hurst and Lyapunov, live
The three shifts we bet on
The three shifts we bet on
6:26 AM. The library, nearly all ours
6:26 AM. The library, nearly all ours
The working structure, phase by phase
The working structure, phase by phase
$160 a drone node, $700 a queen
$160 a drone node, $700 a queen
2:45 AM. Coffee break, then back to it. ☕
2:45 AM. Coffee break, then back to it. ☕
Queens and drones — the Hive
Queens and drones — the Hive
Who did what
Who did what

[08]The certificate

Certificate for AI for Sustainability Hackathon 2026Delivered ✓

Certificate found. 📬

Unlike HackNexus, this one actually arrived. 📬✅ Certificate of participation, team T144, Evolve AI, all six names on it. It's framed here as proof that certificates do exist. 🖼️😌

“The first time we went properly deep into research: state space models, chaos theory, mesh networks and real hardware costs. Half asleep in a library at 4 AM, it stopped feeling like a hackathon and started feeling like research.”

Next hackathon

NeuraCode 1.0: Where Logic Meets Glory

← All hackathons