AI for QA Engineers

Traditional software either works or throws an error. AI systems fail gradually, silently, and non-deterministically — a completely different failure profile requiring a different testing mindset. Hallucination taxonomy: factual hallucinations, faithfulness failures, source fabrication, and confident wrong answers — how to identify and categorise each. Non-determinism: why the same input produces different outputs and how to test probabilistic systems rigorously. Model update failures: how silent model changes break production systems without any visible deployment. RAG failure modes: retrieval failures, chunk boundary problems, context window overflow, re-ranking errors, and context-answer inconsistency. Agent failure modes: tool-use errors, infinite loops, memory contamination, and cascading failures in multi-agent pipelines. Building a failure mode catalogue for your AI system type.

The science and practice of evaluating LLM outputs objectively — moving beyond 'it looks right' to measurable, reproducible quality scores. Reference-based metrics: BLEU, ROUGE, METEOR, BERTScore — when each is appropriate and where each fails. Reference-free evaluation: G-Eval, GPTScore, and LLM-as-judge rubrics for subjective quality dimensions. RAGAS deep dive: faithfulness, answer relevance, context recall, context precision — practical implementation and interpretation. Building custom evaluation rubrics for domain-specific tasks: legal document Q&A, financial data extraction, technical support accuracy. Golden dataset construction: how to build, curate, and version a high-quality evaluation dataset. Statistical significance in LLM evaluation: sample sizes, variance, and avoiding misleading benchmark conclusions. Evaluating multilingual and non-English outputs.

RAG systems have a complex failure surface — retrieval, generation, and the interaction between them. This module covers every layer. Retrieval quality: precision@k, recall@k, MRR, NDCG. Embedding model evaluation on your specific document domain. Chunking strategy testing: validating that different chunking approaches improve or degrade end-to-end answer quality. Hybrid search validation: testing BM25 + dense retrieval combinations for quality-latency trade-offs. Re-ranker testing: cross-encoder precision impact on downstream answer quality. Answer faithfulness testing: detecting hallucination over retrieved text. Context relevance: validating retrieved chunks are genuinely relevant. End-to-end regression: detecting when any component change degrades overall quality. Load and latency testing for RAG APIs under realistic query distributions.

AI is transforming test engineering itself. This module covers the theory and practice of using AI to generate, validate, and maintain test suites at scale. Prompt engineering for test generation: writing prompts that produce high-quality, non-redundant test cases. Generating functional tests: unit tests, integration tests, and API contract tests using LLMs. Edge case and boundary generation: systematically exploring input spaces that human testers miss. Adversarial input generation: AI-generated malicious inputs for security and robustness testing. Test data generation: producing realistic synthetic datasets without using production data. Validating AI-generated tests: evaluating and filtering AI-generated tests for quality and coverage. Maintaining AI-generated suites: keeping large AI-generated test suites current as code evolves. AI for test documentation: generating test plans, descriptions, and coverage reports.

Modern AI systems increasingly process images, PDFs, audio, and mixed-media inputs. Testing multimodal systems requires new approaches beyond text-based evaluation. Vision LLM testing: evaluating vision-capable models from OpenAI, Anthropic and Google on image understanding tasks. Accuracy metrics for visual question answering, image classification, and visual document understanding. Document extraction testing: validating structured data extraction from PDFs, scanned documents, forms, and tables. OCR pipeline testing: character error rate, word error rate, and semantic accuracy for scanned content. Layout understanding validation: table extraction, column detection, and document structure parsing accuracy. Multi-language document testing: extraction accuracy across different scripts and mixed-language documents. Regression testing for document AI: detecting when model updates change extraction behaviour. Testing agentic multimodal pipelines that chain vision LLMs with text models and structured output parsers.

AI systems have a fundamentally different attack surface from traditional software — and QA engineers are increasingly responsible for security validation. Prompt injection testing: designing direct injection attacks, indirect injection via document corpora (a critical RAG vulnerability), and multi-hop injection chains. Measuring injection success rate and validating defences. Jailbreaking and model bypass: adversarial prompts, roleplay exploits, many-shot jailbreaking, and encoding attacks. PII leakage testing: probing for training data extraction, sensitive information disclosure, and context pollution in multi-tenant systems. Insecure output testing: XSS via LLM-generated HTML, SQL injection via LLM-generated queries, command injection via agentic tool calls. Denial of service: inputs causing excessive token consumption, infinite loops, or resource exhaustion. OWASP LLM Top 10 structured test coverage with pass/fail criteria. Building an AI security regression suite.

Regulated industries require AI governance controls that actually work — and QA engineers are increasingly responsible for validating them. Explainability testing: validating that SHAP values, LIME explanations, and attention visualisations accurately reflect model behaviour. Explanation stability: do explanations change meaningfully for similar inputs? Bias and fairness testing: demographic parity, equalised odds, and calibration across protected groups — designing bias test datasets and interpreting results. Audit trail completeness: validating every AI decision is logged with full context, inputs, outputs, and timestamps. Immutability testing: verifying audit logs cannot be altered. Human override testing: validating override mechanisms work correctly and cannot be bypassed. Model card validation: testing that model cards accurately reflect actual behaviour on held-out evaluation sets.

AI systems can degrade silently after model updates, prompt changes, data drift, or infrastructure changes. This module covers systematic regression testing and CI/CD integration for AI systems. Regression strategy: what to test after a model update, prompt change, retrieval index change, or infrastructure change. Golden set regression: maintaining a curated regression dataset running automatically on every change. Snapshot testing for AI: storing expected outputs and alerting on deviation. Differential testing: running two model versions simultaneously and flagging significant quality differences. CI/CD integration: adding AI evaluation gates to GitHub Actions, GitLab CI, or Jenkins. Quality thresholds and deployment gates: defining passing scores and blocking deploys that fail. Rollback strategies: detecting production quality regressions and automatically rolling back to the previous model version.

Testing before deployment is necessary but not sufficient. Production AI systems drift, degrade, and behave differently than staging environments. LLM-specific observability: what to log, what to measure, and how to store LLM interaction data at scale. Real-time quality dashboards: latency SLOs, token cost, error rates, hallucination rates, and user satisfaction signals. Input distribution monitoring: detecting when user query patterns shift — a leading indicator of quality degradation. Output quality monitoring: sampling-based evaluation, automated quality scoring, and anomaly detection. User feedback loops: explicit ratings, implicit signals (follow-up questions, clarifications), and session-level satisfaction. Drift detection: statistical methods for data drift, concept drift, and model performance drift. Alerting strategies and on-call for AI systems. A/B testing in production: statistical design for LLM comparison experiments.

The capstone integrates all 9 modules into a complete AI quality engineering artefact — demonstrated to a panel of HYVE engineers and industry guests at Demo Day. Each participant selects an AI system and delivers: a comprehensive test strategy document covering all failure modes and test coverage matrix; a working evaluation framework with RAGAS metrics, custom rubrics, and golden dataset; a security test report from a full adversarial campaign; a governance compliance checklist with validation evidence; a CI/CD-integrated regression pipeline; a production monitoring dashboard with alerting; and a live demonstration with panel Q&A. Capstone systems are assessed across coverage, rigour, automation quality, documentation, and the quality of insights generated.

Learning Outcomes

• ✓Design and implement comprehensive LLM evaluation frameworks using RAGAS, DeepEval, and LLM-as-judge — measuring faithfulness, answer relevance, context precision, and custom domain-specific quality rubrics
• ✓Build production RAG test suites covering retrieval quality (precision, recall, MRR), answer faithfulness, context relevance, and end-to-end pipeline accuracy across diverse document corpora
• ✓Use AI to generate, maintain, and evolve large test suites — including edge cases, boundary conditions, and adversarial inputs that human testers routinely miss at scale
• ✓Build AI regression testing pipelines that automatically detect when model updates, prompt changes, or data drift silently degrade system quality before it reaches end users
• ✓Test multimodal AI systems — validating document extraction accuracy, structured data extraction from scanned sources, image classification reliability, and vision LLM output quality
• ✓Perform adversarial and security testing on AI systems: prompt injection attacks, jailbreak probing, PII leakage testing, indirect injection via document upload, and OWASP LLM Top 10 validation
• ✓Test AI governance and compliance controls: validate explainability outputs, audit trail completeness, bias metrics across demographic groups, and human override mechanism correctness
• ✓Build production quality monitoring systems with real-time dashboards, statistical drift detection, automated alerting, and A/B testing frameworks for rigorous LLM comparison

FAQs