PULSAR Platform

From raw logs into a system you can ask

The AI-native platform for teams integrating software-defined vehicles — the whole system, or the part you own. PULSAR reconstructs how each software version actually behaves, from the runtime data you already record.

PULSAR · Overview

PULSAR / Overview

software 7.4 ▾ 96 runs last 24 h Ask PULSAR
142control units
3,412software components
86,900interactions

in one software state of one program

The reconstructed system model of a vehicle programme: four labelled domains — perception and ADAS, cockpit and connectivity, body and comfort, powertrain and chassis — each holding six named control units, all hanging off a labelled vehicle backbone of four zonal and compute nodes. Live traffic runs along the links between them. Illustrative, synthetic data.PERCEPTION & ADASFront CameraRadar FusionLidar ControlSurround ViewSensor FusionObject TrackerCOCKPIT & CONNECTIVITYHMI ClusterHead UnitVoice AssistNav ServiceTelematicsOTA AgentBODY & COMFORTComfort ECUDoor ControlHVAC ControlSeat ControlAmbient LightAccess ControlPOWERTRAIN & CHASSISPtrain GatewayCharge ControlBattery MgmtBrake ControlSteering CtrlThermal MgmtVEHICLE BACKBONE · AUTOMOTIVE ETHERNET + CAN-FDZonal Ctrl FLHPC ComputeZonal Ctrl FRVehicle GatewaySOME/IPETHCAN-FD
Ask across software 7.4 — 96 runsAsk
  • Where does 7.4 deviate from learned normal?
  • What behaviour changed between 7.3 and 7.4?
  • Which component started this failure?
Figure: a PULSAR Overview screen. The reconstructed system model of a vehicle program fills the frame — four labelled domains, twenty-eight named control units, and a labelled vehicle backbone, with live traffic running the links between them. Underneath, the questions the platform can be asked across it. Illustrative synthetic data.

Where we built it

Started in automotive. Built for complex systems.

We start in automotive integration. It is the domain we have lived in. How we came here →

Robotics and medical devices are next — same problem, different domain. PULSAR reads logs, not vehicles.

Why automotive is the hard case

Our first target complex industry — and where the problem hurts most.

150+

control units in one car, from different suppliers.

100M

lines of code across them. Nobody holds the whole picture.

Strong IP protection: across every company boundary, software arrives as a black box — in both directions.

The platform

Watch the platform being built

One figure, assembled as you scroll. Five chapters: your runs go in and the model builds itself, then three modules read it.

PULSAR · Overview

PULSAR / Overview

software 7.4 ▾ 96 runs Ask PULSAR Upload runs
PERCEPTION & ADASFront CameraRadar FusionLidar ControlSurround ViewSensor FusionObject TrackerCOCKPIT & CONNECTIVITYHMI ClusterHead UnitVoice AssistNav ServiceTelematicsOTA AgentBODY & COMFORTComfort ECUDoor ControlHVAC ControlSeat ControlAmbient LightAccess ControlPOWERTRAIN & CHASSISPtrain GatewayCharge ControlBattery MgmtBrake ControlSteering CtrlThermal MgmtVEHICLE BACKBONE · AUTOMOTIVE ETHERNET + CAN-FDZonal Ctrl FLHPC ComputeZonal Ctrl FRVehicle GatewaySOME/IPETHCAN-FDFront CameraLidar ControlSensor FusionZonal Ctrl FLOTA AgentVehicle GatewayCharge ControlWRONG ORDERTOOK TOO LONG · 210 msFront CameraZonal Ctrl FLComfort ECUTWO SOFTWARE STATES, HELD AGAINST EACH OTHERSTATE 7.3STATE 7.4Front CameraMESSAGE ORDERA → B → CA → C → BZonal Ctrl FLTRANSITION TIME120 ms300 msComfort ECUCPU LOAD71 %94 %HMI ClusterZonal Ctrl FRVehicle GatewayCharge ControlSYMPTOMTIMING 11× LONGERCPU LOAD 71 → 94 %ROOT · CHARGE CONTROL
Ask across software 7.4 — 96 runsAsk
  • Where does 7.4 deviate from learned normal?
  • What behaviour changed between 7.3 and 7.4?
  • Which component started this failure?
05 / 05 the whole platform, and an agent you can ask across it

PULSAR Engine · before any module can answer anything

Upload your runs — the model builds itself

One screen, one continuous action: as the upload progresses, components appear one by one on the canvas below — and once it completes, the finished model is ready, with behavior fully reconstructed.

Inside the engine

Anomaly Detection · an integration engineer, on release day

What is not normal in this run?

Every run is measured against a baseline PULSAR learned from your program’s own history. Two things go wrong on a propagation path: steps happen in the wrong order, or a step takes too long to transition. Both are marked on the path.

How normal is learned

Root-Cause Analysis · whoever is on the hook for the escalation

Which component started it?

Describe the symptom in your own words. PULSAR traces the same paths backwards to the component where the chain breaks — and shows exactly why.

Root-Cause Analysis

Change Detection · a release lead, two weeks after a green build

What changed between two software states?

Two states of the same program, held side by side. It’s a diff of what the system did, not of the code that changed — and every difference tells you where it happened and what kind it is: message order, transition time, or a measurement that moved.

Ask PULSAR · the same platform, in your own words

What else can I ask?

Everything the screens show, you can also just ask for — and more. Bring your own case and explore it in your own words, against your own model.

What we are building next

The same engine. More modules on it

In development · not in production

Agent Builder

Turn a question your team asks every week into a reusable agent, grounded in the same model the existing modules use.

An invitation

Your custom module

If your program asks a question the existing modules do not answer, the model already has the answer. Just tell us what it is.

Discuss it with us →

A model per software state

One vehicle. A different system version every week. A current model for each of them.

Every release is a combination of component versions that has never run together before. PULSAR does not keep only one model and update it — it re-derives the model from runtime for each system version you run.

KW 032026 State 41
  • ADAST1 Arden4.2.1
  • ChassisT1 Boreas3.1.7
  • PowertrainT2 Dorado2.4.1
baseline
KW 072026 State 44
  • ADAST1 Arden4.3.0
  • ChassisT1 Boreas3.2.0
  • PowertrainT2 Dorado2.4.1
2 of 3 moved since KW 03
KW 102026 State 47
  • ADAST1 Arden4.3.0
  • ChassisT1 Boreas3.4.1
  • PowertrainT2 Dorado2.6.0
2 of 3 moved since KW 07
KW 142026 State 51
  • ADAST1 Arden4.5.2
  • ChassisT1 Boreas3.4.1
  • PowertrainT2 Dorado2.6.3
2 of 3 moved since KW 10
KW 182026 State 55
  • ADAST1 Arden4.5.2
  • ChassisT1 Boreas3.6.0
  • PowertrainT2 Dorado2.7.1
2 of 3 moved since KW 14
KW 222026 State 58
  • ADAST1 Arden4.6.0
  • ChassisT1 Boreas3.6.0
  • PowertrainT2 Dorado2.7.1
1 of 3 moved since KW 18
KW 272026 State 62
  • ADAST1 Arden4.6.1
  • ChassisT1 Boreas3.7.2
  • PowertrainT2 Dorado2.8.0
3 of 3 moved since KW 22
KW 312026 State 66
  • ADAST1 Arden4.6.1
  • ChassisT1 Boreas3.7.2
  • PowertrainT2 Dorado2.8.4
1 of 3 moved since KW 27
KW 362026 State 70
  • ADAST1 Arden4.8.0
  • ChassisT1 Boreas3.9.0
  • PowertrainT2 Dorado2.8.4
2 of 3 moved since KW 31

model rebuilt · state 41 · current

Three successive software states of the same vehicle, cycling. For each one, PULSAR rebuilds the dependency model from that state’s own runtime data. Illustrative, synthetic data: the supplier codenames are constellations, not sourcing relationships.

Get started

Bring us your recorded runs

We build a model of your system. You pick the modules your team needs.

You bring

Recorded runs from the benches you already have — a full vehicle, one domain, or a single component

Whatever they emit, in whatever format. Plus part-time access to one engineer who knows the program. No new tooling on your benches, and you write nothing.

  • MDF4 / MF4
  • BLF · ASC (CAN)
  • DLT
  • OTLP
  • ROS bags
  • plain text logs
  • CSV
  • … or yours

We build

4–6 weeks
any format

From your runs to your first answer

Whatever your benches emit, we write the adapter for your format.