Synthetic Sensor Data Qualification Expert

The mandate

The developer uses simulated camera, lidar and thermal scenes to train detection of track intrusion, displaced equipment and weather-obscured defects that field datasets rarely capture. Its product council cannot determine whether improved test scores reflect useful scenario expansion or synthetic signatures the model exploits, and the narrow problem is to qualify permissible use before a new corridor pilot.

The deliverable is a Synthetic Sensor Data Qualification Dossier containing an operating-domain taxonomy, source-to-scene parameter map, sensor-fidelity tests, scenario coverage measure, artefact hunt, mixing and weighting rules, domain-gap experiments, rare-hazard performance analysis and a release control for future generator changes. It must state where synthetic evidence is unsuitable rather than forcing one portfolio-wide answer.

Work begins 18 January 2027. Milestone one on 29 January is the agreed domain map, evidence inventory and qualification protocol; milestone two on 26 February provides generator diagnostics, controlled mixture experiments and field-comparison findings; milestone three on 26 March delivers the signed dossier, witnessed rerun, permitted-use matrix and owned remediation backlog.

Acceptance requires the perception lead to reproduce every reported experiment from retained scene parameters, seeds and model configurations, while the assurance chair traces each approved use to a measured field or controlled-test consequence. Results must survive a held-out real-world sample and an artefact-focused challenge; a prettier simulation or aggregate accuracy gain cannot pass the engagement.

The client will provide generator access, sensor specifications, calibration records, raw field sequences, labelled incidents, model checkpoints, controlled track time and named simulation and safety engineers. The scope excludes rail-operation certification, collection of new public-track data, production model remediation and changes to vehicle hardware after the accepted backlog transfers.

Why this is external work

Simulation engineers are accountable for generator adoption, while perception teams need more rare events and therefore cannot independently arbitrate the domain gap. Product assurance understands field risk but lacks specialist methods for diagnosing learned synthetic artefacts. An external qualification creates one falsifiable evidence chain before synthetic volume becomes an unquestioned proxy for coverage.

What you will own

• Define the operating-domain and hazard cells that synthetic data claims to extend, including sensor state, weather, infrastructure, object and motion conditions.
• Trace geometry, material, illumination, weather, calibration, noise and rendering parameters from generator configuration into each training and evaluation scene.
• Design artefact probes that test backgrounds, edges, textures, missing noise, metadata and rendering shortcuts a perception model could exploit.
• Compare real-only, synthetic-only and controlled-mixture training through held-out field data, rare-hazard recall, calibration, false alarms and failure clustering.
• Quantify coverage contribution without counting near-identical generated variants as independent scenarios or treating parameter range as realistic frequency.
• Decide permitted, conditional and prohibited synthetic uses for pretraining, augmentation, stress testing, validation and safety-case support.
• Transfer the runnable experiment suite, scene manifest, domain-gap register, change trigger and qualification decision log to internal assurance owners.

Candidate qualifications

• Qualified synthetic camera, lidar, radar or thermal data for an autonomous, robotic or safety-relevant perception system.
• Exposed a generator artefact or simulation shortcut that improved benchmark performance while reducing transfer to real operating conditions.
• Built controlled dataset-mixture experiments with stable checkpoints, seeds and field holdouts that separated data value from training variance.
• Assessed rare-event and calibration performance beyond aggregate precision and recall across a meaningful operating-domain taxonomy.
• Linked sensor physics, simulation parameters and learned behaviour closely enough to challenge both rendering and perception specialists.
• Delivered a reusable synthetic-data release control whose owners could reassess generator changes after the consulting engagement ended.

Non-negotiables

• The named expert must work in Prague during protocol design and acceptance and attend all three Brno proving-ground days.
• No rail footage, generator asset, calibration record, checkpoint or scene parameter may leave the controlled client environment.
• Will disclose commercial or research ties to simulation engines, sensor vendors and autonomy platforms before accessing comparative evidence.
• Must state a prohibited-use conclusion where field transfer is not demonstrated, even if synthetic production capacity is already contracted.

1. 49 words maximum. Describe one synthetic sensor artefact your model learned and the field test that exposed it.
2. 49 words maximum. How would you prove that more generated rare-hazard scenes add coverage rather than correlated visual variants?
3. 49 words maximum. Confirm ten-week capacity, Prague and Brno attendance, and all simulation or sensor relationships relevant to independence.