Four operations, in full.

Situation

Label requests arrived ad hoc from multiple feature teams with no shared prioritisation. Volume requirements for European perception data were climbing faster than the operation could absorb, and quality had to hold near 99% while they did.

Scale

• Five perception modules: Speed Limits, Traffic Signs, Blockage, 3D Dynamic Objects, Vulnerable Road Users
• Nine European markets, Germany as primary focus
• Multiple labelling suppliers plus internal revision teams
• 2D and 3D annotation running in parallel

What I owned

• Owned the delivery roadmap and operating model for ADAS training and validation data
• Restructured scattered label requests into geo-segmented campaigns organised by country cluster
• Translated feature-team requirements into campaign configurations, annotation guidelines and acceptance criteria
• Led label strategy across 3D annotation programmes, including ontology design and quality frameworks
• Configured and calibrated auto-labelers across perception modules to reduce manual annotation load

Operating model

• Campaign planning replaces request queues
• Acceptance criteria defined before annotation starts, not after
• Machine first pass, human verification on what the model is unsure about
• Executive reporting on throughput, quality, risk and cost drivers

Situation

Quality assurance for 3D dynamic objects and vulnerable road users depended on vendor-side tooling and vendor-side reporting. Depth and occlusion errors are close to invisible in a summary report and tend to surface much later, as model behaviour.

Scale

• Multi-vendor annotation with separate revision teams
• Ambiguous sign properties spanning prohibited and danger classes
• 3D and 2D modules with different failure modes

What I owned

• Defined ontology, annotation guidelines and acceptance criteria for dynamic object and VRU modules
• Implemented and validated configuration changes for prohibited and danger-sign properties
• Coordinated cross-functional testing and controlled rollout across labelling and revision workflows
• Drove defect triage into a structured loop connecting annotation, model feedback and simulation

Operating model

• Ontology and guidelines treated as product artefacts, versioned and owned
• Configuration changes tested and rolled out deliberately rather than announced
• Revision workflow treated as part of the pipeline, not an exception path

Situation

Merging two national operators means merging two sets of enterprise processes without customers noticing. Transition risk sat across engineering, operations, commercial and customer success simultaneously.

Scale

• 5+ enterprise processes in the Vodafone and Three UK merger
• EMEA and APAC stakeholders
• AWS, Azure and GCP suppliers under one platform estate

What I owned

• Directed end-to-end service transition across the affected enterprise processes
• Governed hyperscale suppliers through SLAs, quarterly business reviews and performance scorecards
• Standardised rollout blueprints and readiness assessments
• Introduced OKR-linked dashboards and dependency tracking

Operating model

• Readiness assessed against a standard before any rollout
• Supplier performance measured continuously, not reviewed annually
• Dependencies tracked as first-class objects

Situation

Reviewing 3D annotation through a vendor's interface meant the team could see a quality number but not the failure. Verification depended on somebody else's tooling and somebody else's definition of correct.

Scale

• 3D dynamic object and VRU labelling at programme volume
• Multiple annotation platforms already in the estate

What I owned

• Leading development of a production-grade 3D annotation viewer for Q4 2026 release
• Translated recurring operational failures into product requirements for in-house visualisation and QA
• Specified what reviewers need to see to judge depth and occlusion, rather than replicating vendor views

Operating model

• Operations insight becomes a requirement, not a complaint
• Quality assurance brought in-house where the failure mode is invisible in reporting