- Own the acceptance test plan: define how 95% detection accuracy and - 5% area variance are measured, and enforce the split between the calibration set used for tuning and the held-out set used for acceptance

- Build and maintain the coverage matrix across design types and symbol classes, flagging thin or empty cells as known gaps

- Report accuracy per symbol class with confidence intervals, not just an aggregate figure, and enforce the per-class minimum instance count below which a class is reported as provisional

- Run no-regression checks across fine-tune rounds so a new model cannot pass on average while regressing on a rare but critical class

- Build the Knowledge Bot evaluation suite: a curated Q&A set with known-correct answers, scored for groundedness and faithfulness against every prompt and model version

- Perform adversarial and AI security testing before production and after material architectural change prompt injection, jailbreak attempts, RAG-specific attacks, cross-user data isolation and AI data leakage

- Validate correction-quality controls: spot-audit ambiguous classes, track inter-reviewer agreement, and confirm conflicting human corrections are adjudicated before entering the labelled set

- Stand up drift monitoring and produce the periodic model-health report covering drift indicators, correction rates, confidence-score trends, retraining events and acceptance results by version

- Run functional and integration testing of the HE2 workflow alongside the model validation work

What you need :

- Python, and the ability to write your own evaluation harnesses rather than only driving a test tool

- Solid grasp of ML evaluation: precision, recall, per-class metrics, confidence intervals, sampling bias, and why a held-out set must stay held out

- Experience testing an ML or AI system in production - you have caught a model problem, not only an application bug

- Test automation and CI, with pytest or equivalent

- API testing and the discipline to build reproducible, versioned test datasets

- Clear written reporting: your reports go to the customer and support a payment gate

Nice to have :

- LLM red-teaming or prompt injection testing experience

- Familiarity with RAG evaluation frameworks and groundedness scoring

- Statistical background: clustering effects, effective sample size, stratified sampling