Work

What we've already built, in real conditions.

A few engagements run for teams that had valuable data and a concrete problem to solve.

Security · insider threat detection

From scattered signals to a trainable detection system.

The need

Spotting risky behaviour coming from inside, where static rules mostly produce noise and no labelled dataset existed to learn from.

What we built

We built the dataset end to end (collection, structuring, labelling, documentation), then the classification system trained on it.

The asset they kept

A labelled, documented dataset they can reuse to retrain, compare or audit models, and a classification system running on their own data.

What it means for a director: A detection capability that learns from your own data, instead of rules your team spends its time tuning.

Beauty · evaluating a live system

Evaluating a product assistant with synthetic conversations.

The need

Knowing whether the assistant actually answered correctly for real customers, beyond a handful of manual tests that covered neither the range of profiles nor multi-turn exchanges.

What we built

Persona-driven multi-turn conversation generation, played against the live RAG, with the criteria needed to judge each answer.

The asset they kept

A reproducible evaluation harness: personas, conversations and criteria replayable on every change to the system, so regressions surface before customers find them.

What it means for a director: Measured proof the assistant answers correctly before your customers see it, and a test you can replay on every update.

AI R&D · model behaviour

Understanding what an agent really does, from its traces.

The need

Execution traces piled up without ever becoming usable: model behaviour stayed an impression, argued anecdote by anecdote.

What we built

A pipeline that turns agent execution traces into a structured, analysable dataset.

The asset they kept

A dataset of how their agents actually behave: a basis for analysis, comparison and continuous improvement rather than a stream of logs.

What it means for a director: You finally know what your AI does day to day: measurable facts for the board, not impressions.

Production AI · model deprecation

Switching language models without starting over.

The need

Providers deprecate their models: the one the system was built on disappears, the migration comes with a deadline attached, and nobody knows what the replacement will break.

What we built

A regression test suite built from their own use cases: every candidate model is replayed on the same tasks and compared against the reference behaviour, answer by answer.

The asset they kept

A non-regression suite replayable at every deprecation notice or new model release, turning model comparison into a measurement rather than a feeling.

What it means for a director: You change models when you decide to, with evidence that nothing degraded, instead of being held to your provider's schedule.

AI R&D · neuro-symbolic pipelines

Pipelines that let models learn from themselves.

The need

Improving a language model on a specific domain without depending on human annotation at every iteration.

What we built

Neuro-symbolic self-distillation pipelines: the model generates its own examples, the symbolic layer verifies them and filters out what doesn't hold, and only what passes is used for training.

The asset they kept

A reproducible pipeline and the data it produces, re-runnable with every new version of the model.

What it means for a director: A model that improves on your domain without tying up your experts every iteration, and without sending your data elsewhere.

The common thread

In every one of these engagements the deliverable isn't a demo: it's a documented, reproducible data asset the team keeps and reuses. That's exactly what the assessment is there to scope for you.

Your turn

Let's talk about what your data could produce.

Thirty minutes, free and with no obligation, to find out whether an assessment is the right next step.

Book an AI-readiness assessment