Click into the working systems.

What this demonstrates

One strategic question can be transformed into a structured decision package: hypotheses, evidence references, actions and validation tasks in one typed workflow.

When this pattern is useful

Executive questions are answered ad-hoc, and the reasoning behind the answer cannot be reviewed later.

Where the claim stops

Core company and market data is synthetic. No causal attribution, no measured executive time saving, no commercial impact claim.

What this demonstrates

Retrieval-augmented answers over unstructured review text, with the retrieved evidence kept visible alongside the generated answer.

When this pattern is useful

A team needs answers out of documents, tickets or reviews and cannot accept unsourced output.

Where the claim stops

Demo data and demo scale. Answer quality on your own corpus is an open question until tested.

What this demonstrates

A structured product built on retrieval and event tracking — practice tracks, explanations and progress — rather than a chat window over a model.

When this pattern is useful

Internal enablement or knowledge work needs structure, progress and explanations that can be audited.

Where the claim stops

An internal product, not a client deployment. No learning-outcome or retention claims.

What this demonstrates

3 run modes, 2 export formats and 1 CRM handoff contract, all verifiable in the repository. An agent workflow can keep its evidence and hand off cleanly to an operational system.

When this pattern is useful

Manual research work is repetitive, and the output has to arrive in a system of record with its evidence attached.

Where the claim stops

A capability demo. No client pipeline volumes, conversion rates or commercial outcomes.

What this demonstrates

3 ML training paths, 4 worker roles and 1 ML feature contract defined in the repository — the scaffolding a scoring system needs before any model is trusted.

When this pattern is useful

Inbound volume needs triage and routing, and the scoring logic has to stay reviewable as it changes.

Where the claim stops

Under development. The ~€850k uplift referenced elsewhere is modelled in this workflow, not realised revenue.

What this demonstrates

12 deterministic research topics, 25 defined evaluation cases, and 0 paid inference APIs required at runtime. Retrieval quality can be evaluated on a defined case set, on local infrastructure.

When this pattern is useful

Sensitive material cannot leave your infrastructure, or inference cost per query has to stay near zero.

Where the claim stops

Evaluation results are not yet published. Defined cases are not measured benchmark outcomes.