AI Engineering for Automotive

Perception evidence your validation team can read, re-run and challenge: scenarios, sensor timing and safety-case inputs.

Start a conversation

A perception model that scores well on a benchmark can still fail on the road. The rain scenario was never in the test set, sensor timestamps drift at highway speed, or a software update ships without a regression gate. Each failure gets fixed once and returns later.

The fix is engineering evidence around the model, not a different model. We work with perception, validation and safety teams on civilian vehicle programmes and the automotive supply chain.

Concept car driving through a green city

Where we help

Three Questions Automotive Teams Bring Us

  • Scenario evidence

    Scenarios

    Which situations must be covered, and does the model hold in them?

  • Sensor timing

    Timing

    Calibration and time synchronisation that fusion can rely on.

  • Safety evidence

    Safety

    Inputs your safety case can use, and an audit trail behind them.

Scenarios

Which Scenarios Must We Cover, and Does the Model Hold up in Them?

Before a perception model is released, you need to show which lane, weather, lighting and traffic situations it has been tested in, and why those are enough for the operating domain. We help turn fleet logs and near-misses into prioritised scenarios, gate test slots on whether they conform to the operating domain, and set up simulation and replay that confirm a fix stays fixed.

Scenario evidence, in depth

Sensor timing

Is the Fault in the Model, or in Timing and Calibration?

Sensor fusion is only as good as the clocks and calibration beneath it. A timing spike on one bus can look like a perception fault, and a calibration that went stale looks like a model fault. We work on finding where the jitter comes from, setting a defensible timing budget, and keeping calibration honest with staleness checks that run without anyone remembering to.

Timing and calibration, in depth

Safety evidence

What Evidence Does Our Safety Case Need from the Perception Model?

Functional-safety and SOTIF work asks for evidence an assessor can follow: a defined item, a hazard analysis, scenarios tied to safety goals, and a record of what was tested and what passed. We help perception teams produce that evidence in a form the safety roles can read and challenge. The safety case itself, and its approval, stay with them.

Safety-case inputs, in depth

Further reading

Where Our Automotive Thinking Is Written Down

Our articles on safety integrity levels, what a robustness audit covers, and moving perception models onto embedded hardware.

AI Is Reshaping the Automotive Industry

Automotive edge AI in the car and on the line: what latency, power and INT8 portability limit, and what to check before trusting a perception model.

2019
Founded in Budapest
10+
Patents co-authored with clients
1
Client per niche

Client Testimonials

Automotive AI Engineering FAQ

Which Scenarios Must We Cover Before We Can Say the ODD Is Addressed?

The ones your operating domain makes reachable, taxonomy first: lane markings, weather, lighting and occlusion, each tied to the hazard it exercises. We build the scenario set and coverage evidence with your team, through Production AI Reliability. Read more: Building a Defensible Scenario Taxonomy and Safety Case.

Can We Find Which Sensor Bus or Clock Is Causing Jitter Spikes at Highway Speed?

Usually yes: we trace timestamps across the bus, grandmaster and fusion stages to isolate the source, then set a jitter budget the fusion can live with. The timing checks stay with your team to re-run. Read more: PTP Time Sync Jitter at Highway Speed in Sensor Fusion.

Can We Replay the Log That Triggered a Failure and Confirm the Patch Fixed It?

Yes. We replay the logged event through the old and patched pipelines and compare the outputs, then keep the case in your regression suite so it cannot return unseen. Read more: Proving Sensor Fusion and Alerting Fixes With Logged-Data Replay.

What Evidence Will a Functional-Safety or SOTIF Assessor Expect to See?

That is the assessor's call, but they read evidence that is traceable: scenarios, results, limits and decisions linked to the hazards. We structure our outputs so your safety team can place them in the case. Read more: Building an ADAS Validation Evidence Package for Safety Sign-off.

Where Should We Start?

Start where the decision is closest. If a perception model is heading for a release gate, Production AI Reliability covers the scenario evidence and robustness checks. If a model is already running and you need to keep it honest across software and hardware changes, the Production AI Monitoring Harness is the entry point. If you are not sure, tell us about the programme and we will route you.

How we work with automotive teams

An engagement has a fixed scope and ends in something your team keeps and can re-run. We work alongside your safety, validation and test functions; we do not substitute for them.

Lands in the Production AI Monitoring Harness: 4–10 weeks, milestone or fixed-price.

Heading into a release gate, a sensor change or an assessor review? Tell us the programme and we will route you to the right starting point.

Start a conversation
Concept car driving through a green city

Keystone Deep Dives

Show all 105 deep dives