Evaluation and data work for AI systems
We build the measurement layer: annotated datasets, agent evaluations, purpose-built benchmarks and leaderboard runs — for teams who need to know what their systems actually do, not what a demo suggests.
What we do
Data annotation
Labelled datasets built to a written spec, with the guidelines, the adjudication process and the inter-annotator agreement reported alongside the data.
Agentic evaluation
Measuring how agents behave over multi-step tasks — the harness, the task set and the rubric, plus the trajectory-level analysis that explains a score rather than just reporting it.
Benchmark creation
New benchmarks built from scratch: task design, item writing, difficulty calibration, contamination checks, and a baseline suite so a number means something on day one.
Leaderboard evaluation
Running models against a fixed suite on a repeatable cadence, with versioned prompts and seeds so results stay comparable across runs and across models.
AI-driven mobile apps
iOS and Android products with a model at the centre — the inference path, the latency and cost budget, and the evaluation loop that keeps quality from drifting after launch.
Working with us
Engagements usually start with a scoping conversation about what you are trying to measure and why the current signal is not enough. Tell us what you are building and we will reply from a real address.