banana split Talk to the team

Distributed intelligence for sensor networks

Intelligence, everywhere.

Information is distributed. Shouldn't intelligence be?

Explore the system For developers →
2nd Prize · Singapore Defence Tech Hackathon · Incubated at NUS Defence Venture Labs Sense · Understand · Sign · Synchronise · Infer

01The world and the problem

Every sensor sees part of the picture. The thinking happens in one room.

Feeds travel from every sensor to a central ops room. People there reconcile them into one picture, then send action back out. Everything waits on that room.

  1. 01Sensors capture
  2. 02Feeds travel to the ops room
  3. 03People reconcile
  4. 04Action goes back out
Illustration · Centralised model

One room on the critical path

Every decision waits for the feeds to arrive and be reconciled in one place.

Raw feeds, not understanding

Sensors send what they capture. Meaning is made somewhere else, later.

One link, one dependency

When the link to the room fails, the picture fails with it.

02The split

Coordination without a coordinator

  1. 01

    HQ rises

    Every link runs to one room.

  2. 02

    Congestion

    Feeds pile up. Everything waits on the room, and one link jams.

  3. 03

    The seam

    A hairline of light opens down the room.

  4. 04

    Split and sink

    The room sinks. Queued work returns to its authors' Local Ledgers.

  5. 05

    The architecture

    Peers link directly. Every node holds the same Global Ledger, and infers from it.

Illustration · Centralised model

Information is distributed. Intelligence should be too.

03Every node runs

Five steps. On every node.

  1. Step 01

    Sense

    Each node taps the sensors on the platform it rides.

  2. Step 02

    Understand

    An edge vision-language model turns what the sensor sees into a structured observation.

  3. Step 03

    Sign and append

    Every record is signed by the node that made it, and appended to its Local Ledger.

  4. Step 04

    Synchronise

    Nodes exchange and synchronise signed records. The network keeps a Global Ledger.

  5. Step 05

    Infer

    Edge language models infer across the network, and a shared world state takes shape.

N1 · Camera node Edge compute · online

Sense sensor feed

Understand edge VLM

object
vehicle
motion
heading north-east
source
frame 0412 · N1

Local Ledger signed, appended

  1. 0412N1 · vehicle, NE
  2. 0411N4 · no change✓
  3. 0410N2 · area clear✓

Synchronise peer to peer

  • N2copy received✓
  • N3copy received✓
  • N4copy received✓
  • N5copy received✓
  • N6copy received✓

Global Ledger · committed ✓

Infer edge LM

› new record from N1: a vehicle, heading north-east

› N4 holds a record from the next sector

› shared world state updated on every node

Illustration · Not a recorded run

Take out the network. Take out the nodes. Command still holds.

04The architecture

No ops room on the critical path. Every node keeps the record.

01 · Nodes

Distributed sensor nodes

Camera nodes, drone nodes and ground sensors, with acoustic sensors as a fourth platform type.

02 · Edge compute

Intelligence at the edge

Edge vision-language models understand what a sensor observes. Edge language models run inference.

03 · Records

Signed at the source

Every observation a node makes is signed by that node.

04 · Local Ledger

Each node keeps the record

Each node keeps a Local Ledger of signed records.

05 · Synchronisation

Node to node

Nodes exchange and synchronise signed records directly, with no coordinator in between.

06 · Global Ledger

One shared ledger

The shared ledger of the network's signed records. A shared world state is inferred from it.

05Evidence

What the experiment demonstrated. And what comes next.

Demonstrated

The pipeline above, as the experiment ran it. Inputs: synthetic data.

Illustration

The diagrams on this page: the split, the layers, link loss and node loss.

Planned

The milestones in the progression below.

Inputs: synthetic data. Detector accuracy: not claimed. Link loss and node loss: illustration.

06Progress

Where it has come from. Where it is going.

  1. Recognition

    2nd Prize

    Singapore Defence Tech Hackathon

  2. Recognition

    Incubated

    NUS Defence Venture Labs

  3. Planned

    Datasets

    Access to more complex datasets (4 months)

  4. Planned

    Proof of concept

    Proof of concept (4 months)

  5. Planned

    Ecosystems

    Ingestion into wider ecosystems

Recognition of the team and its work so far. Not an independent technical evaluation, endorsement or approval for deployment.

07Team

The people behind the split.

  • 01

    Frederick Amal Emerson

  • 02

    Hongyi Lin

  • 03

    Patrick Steve Harrison

2nd Prize · Singapore Defence Tech Hackathon · Incubated at NUS Defence Venture Labs

08Contact

Intelligence without a centre.

Decentralised intelligence and coordination · Shared Global Ledger

Talk to the team Explore the documentation