AI fails differently.
We test differently.

KiwiQA tests the full spectrum of AI and machine learning systems including generative AI applications, large language models (LLMs), Agentic AI systems, AI chatbots, RAG (Retrieval-Augmented Generation) pipelines, recommendation engines, computer vision systems, natural language processing models and AI-integrated enterprise applications. We serve clients across healthcare, financial services, legal, government, e-commerce and logistics sectors where AI failure carries serious regulatory or operational consequences. Our K-ASCI framework provides structured testing coverage across all AI system types, from pre-production validation through continuous post-deployment drift monitoring.

KiwiQA measures hallucination rates through a structured evaluation process. We design adversarial prompt sets across known factual domains relevant to the application's use case — finance, healthcare, legal or general knowledge — then run systematic groundedness evaluations comparing model outputs against verified source material. We apply LLM-as-judge scoring frameworks where a separate model evaluates response faithfulness, and calculate hallucination rates at the 95th and 99th percentile. We establish an agreed baseline rate before production sign-off and validate outputs against defined thresholds. For RAG systems, we additionally test retrieval accuracy and citation fidelity.

The EU AI Act is a regulatory framework classifying AI systems by risk level — unacceptable, high, limited and minimal risk. High-risk AI systems deployed in healthcare, finance, employment, education, critical infrastructure and government must meet mandatory conformity requirements before deployment, including bias testing, transparency documentation, human oversight mechanisms, robustness validation and post-market monitoring. KiwiQA's AI testing framework is aligned with EU AI Act Article 9 requirements and provides the testing evidence documentation that conformity assessments demand. For Australian companies exporting to the EU, compliance is required for any high-risk AI touching EU citizens.

A focused AI testing engagement typically takes 4–8 weeks for initial validation coverage, depending on system complexity, data availability and the number of AI components involved. Scope includes model accuracy validation, bias testing across demographic groups, adversarial prompt testing, performance benchmarking and security assessment. For large-scale enterprise AI systems with multiple models and integrations, initial engagements may run 10–12 weeks. Post-deployment monitoring engagements run continuously in production, with monthly reporting. KiwiQA can mobilise an AI testing team within 48 hours for urgent go-live validations where time-to-market is critical.

Prompt injection is a class of attack where malicious input manipulates an AI system's instructions, causing it to bypass safety guardrails, reveal sensitive system prompts, execute unintended actions or leak confidential data. It is the most critical security vulnerability in LLM-powered applications. KiwiQA tests for prompt injection by running structured attack libraries covering direct injection (malicious user input), indirect injection (poisoned external data sources), jailbreak scenarios and multi-turn manipulation attacks. We map all successful injection vectors, validate the effectiveness of mitigations and produce a risk-rated report with reproduction steps. Our test library is continuously updated as new attack patterns emerge.

Yes. KiwiQA applies demographic parity analysis, equal opportunity testing and intersectional fairness scoring across AI outputs. We test model performance across protected attributes including age, gender, ethnicity, disability status, nationality and socioeconomic indicators to identify where accuracy or output quality degrades for specific subgroups. Our methodology includes dataset audit for representation gaps, output distribution analysis across demographic groups and counterfactual fairness testing. All bias testing is aligned with the EU AI Act, OECD AI Principles, IEEE P7003 and applicable anti-discrimination legislation in Australia, the US and UK.