I turn complex AI data operations into production-scale systems.
Autonomous systems · Computer vision · Human-in-the-loop AI · Data operations at scale
Complex operations that work at small volume and break at large one. The technology changes. The problem does not.
Reliable training and validation data at production scale.
Better data infrastructure for machines reading the physical world.
Fragmented operations become measurable production systems.
“Throughput went up. Quality stopped being predictable.”
Usually a QA design problem, not an annotator problem. Review effort is spread evenly instead of following risk.
“Two vendors read the same ontology two different ways.”
Ambiguity in the spec becomes disagreement in the data. The fix is upstream, in guidelines and acceptance criteria.
“The ML team spends more time managing annotation than modelling.”
Annotation has become an operation without an operating model. It needs ownership, cadence and reporting of its own.
“Millions of frames, and no way to say which ones matter.”
Random sampling keeps rare classes rare. Curation has to be deliberate or the dataset grows without getting better.
“QA finds problems after the dataset has already moved downstream.”
Quality gates sit too late. Detection needs to move next to the step that creates the defect.
“Annotation cost is rising faster than model value.”
Cost per frame is the wrong unit. First-pass acceptance and rework rate are what actually move the bill.
“It works at 100k items. It breaks at 1M.”
Everything that depended on someone checking manually stops working. That threshold is a design problem, not a staffing one.
AI teams training computer-vision models
Annotation quality, feedback loops and dataset operations that hold as volume grows.
Autonomous-systems teams
Perception-data campaigns across markets, edge cases and annotation complexity.
AI infrastructure companies
Product requirements turned into scalable human-in-the-loop operating models.
Enterprises building internal AI capability
Governance, suppliers and quality systems designed around production data.
Most operations describe themselves as measured and behave as managed. The tell is whether a model failure can change what gets labelled next without someone filing a ticket.
I am not a foundation-model researcher or a perception-algorithm engineer. My work sits at the layer where data, product, workflow, vendors and quality meet — which is usually where production AI actually breaks.
ONE FRAME. RAIN, GLARE, OCCLUSION. THE STACK SEES NOTHING.
STACK YOUR FILTERS. FIND THE ANOMALY. SAVE THE FRAME.
MOVE THE POINTER THROUGH THE CONSTELLATION
Professional environments are organisations and programmes worked with directly. Platforms are tools governed, configured or evaluated. Neither implies partnership or endorsement.
Eight questions, about ninety seconds. The full read appears immediately — score, maturity stage, bottleneck and the intervention I would start with.
Throughput sets which failure modes are even possible.
A scoped review of the operation that is actually breaking, and the operating model that would hold at the next order of magnitude.
Under NDA · scoped before anything starts
I take a small number of senior conversations each year, where the role owns the data systems behind the model rather than the model alone.