Quality & Knowledge Engineering
Quality isn't a phase. It's a system.
We embed quality into every layer — test automation on traditional software, hallucination detection on AI systems, and the knowledge architecture that makes your AI actually know what it's supposed to know.
Two things most teams underinvest in
Where quality quietly breaks.
First, QA on traditional software — treated as a checkbox at the end of a sprint instead of a discipline woven into delivery. The result is production bugs, regressions, and a QA team always playing catch-up.
Second, knowledge quality in AI systems — teams ship RAG pipelines and agents without asking whether the AI actually has accurate, current, complete information, or whether it's grounding answers in the right context. If you don't engineer the knowledge layer, you ship a system that confidently answers wrong. InWork fixes both.
Three disciplines
One quality system across software, AI, and knowledge.
Software QA
Test & automateTest strategy, manual & automated functional QA, Playwright/Cypress/Selenium e2e, k6/JMeter load, accessibility (WCAG 2.1 AA), and OWASP security testing — wired into CI/CD so failing tests block merge.
AI System Evaluation
Trust the outputGolden-dataset evaluation, hallucination & RAG scoring (RAGAS, DeepEval), prompt-regression harnesses, and compliance red-teaming — because AI outputs are probabilistic, and 'confidently wrong' is a failure mode you have to test for.
Knowledge Engineering
Know the right thingsDomain taxonomy, chunking strategy, metadata schema, hybrid retrieval (vector + keyword + graph), and ongoing maintenance — the difference between a document dump and a knowledge base your AI can reason over.
The knowledge layer
Architected, not dumped — with an evaluation loop that never stops.
AI evaluation pipelines
Catch what traditional QA can't.
AI fails in ways software doesn't. We build the pipelines that catch it before your clients do.
Hallucination detection
Automated scoring on every model/prompt change against a curated golden dataset; flagged before it reaches production.
RAG evaluation
Retrieval precision/recall, chunking and reranker testing, and embedding-model comparison to validate the right context is surfaced.
Prompt regression
Version-controlled prompts, A/B harness against ground truth, automated alerts on accuracy drops, and one-click rollback.
Compliance testing
Adversarial red-teaming and guardrail validation for TCPA, PHI disclosure, and investment-advice boundaries in regulated domains.
Reference knowledge bases
Knowledge layers we've engineered.
Automotive compliance KB
TCPA/FCC/FTC rules, OEM communication requirements, and state-by-state variations indexed and queried by every agent before an outbound action — compliance enforced at the retrieval layer.
Financial intelligence KB
Portfolio data, research, CRM, and email unified into one queryable layer for a US family office — every answer cites its source document and extraction date, with role-based access.
Veterinary knowledge base
Built from veterinary references and breed-specific guidelines, validated at 91% accuracy against a licensed knowledge base, with hard guardrails against clinical diagnosis.
QA as a culture
Quality is in every stage, not a separate sprint.
Discovery — requirements written with testable acceptance criteria.
Architecture — peer review of data model, API contracts, and integration design.
Development — unit tests ship with code; a PR can't merge without passing tests.
Staging — full regression plus AI evaluation run before every deploy.
Production — monitoring, alerting, and statistical AI sampling from day one.
Before you ship a RAG system