Services / AI validation and investigations
For software companies and organisations in Switzerland and abroad: LLM validation for training teams, and post-incident analysis for AI systems and agents. Remote delivery depends on the agreed access and scope.
01 / AI Validator
For teams working on pre-training, fine-tuning or domain adaptation. Objectives, model access and evaluation depth are agreed together.
An evaluation suite combining relevant benchmarks and representative tasks: comprehension, extraction, reasoning or generation, depending on the project.
You receive
Results by task, language and difficulty, using agreed metrics and error analysis. Specialist assessments require reference answers validated by domain experts.
Comparison of the base model, adapted versions and release candidates under explicit, comparable test conditions.
You receive
An analysis of improvements and regressions. Aggregate scores are accompanied by category-level results to make trade-offs visible.
Tests with verifiable answers to examine correctness, consistency, instruction following and behaviour when information is insufficient.
You receive
Classified errors, documented examples and test results. Any automated evaluators are compared against a human-reviewed sample.
Agreed tests covering input variations, adversarial instructions, inappropriate responses and possible disclosure of sensitive information.
You receive
Test scenarios, conditions and outcomes, with issues and priorities. Access to weights and information about training data determine which checks are feasible.
Examination of training/test separation, sample representativeness and possible overlap, within the available material.
You receive
Findings on risks that may distort measurements. The absence of contamination cannot be established without access to the relevant training data.
Rerunning tests on new versions against agreed acceptance thresholds, documenting the model, prompts, generation parameters and environment.
You receive
A validation report covering results, observed variability, limitations and open issues to support the team’s release decision.
Illustrative scenario
An evaluation example, not a client case.
A software company adapts an LLM to a specialist domain. Before release, it wants to compare it with the original model: does it answer domain questions better? Retain general capabilities? Still follow instructions? Introduce new errors?
An agreed suite measures these aspects separately. The report shows where the checkpoint improves, where it regresses and which results need further investigation.
Method and deliverables
Model, version, domain, languages, objectives and access arrangements: an endpoint or an agreed environment.
Test datasets, baseline, metrics, acceptance criteria and execution conditions.
Appropriate tests and repetitions, error classification and human review of relevant cases.
Comparative report, detailed results, protocol and reproducible materials specified in the engagement.
For the initial contact: base model, training approach, domain, languages and evaluation objective. Proprietary weights, datasets and credentials are shared only under agreed arrangements and conditions.
02 / AI incident forensics
Post-incident analysis for businesses, software companies and law firms investigating a problematic response, information disclosure or an unexpected AI operation.
Examination of available prompts, system instructions, conversations and external content used by the system. Investigation of possible manipulative instructions or conflicting directions.
Correlation of responses, tool calls, operation outcomes and logs from the systems involved. An AI statement alone does not establish that an action occurred.
Reconstruction, where documented, of model version, settings, access, permissions and human approvals. Analysis of the conditions that enabled the observed behaviour.
Chronological organisation of available evidence, identification of sources and distinction between observed facts, hypotheses and gaps. Any reproduction is agreed in a controlled environment.
Illustrative scenario, not a client case.
What request did it receive? Which content did it consult? Did it propose or execute the change? With which permissions and approvals? The investigation seeks evidence in the available logs.
Describe the event, timeframe and system involved. Indicate which conversations and logs are available. Before transferring confidential data or credentials, we agree on materials, access and sharing arrangements.