QA Lead- AI Evaluation & Quality

Educational Qualification

- B.Tech./B.E. or M.Tech./M.S./M.Sc. in Computer Science, Data Science, AI/ML, or a related quantitative discipline (Must have)
- Advanced degree (M.Tech./M.S./Ph.D.) or published/applied work in ML evaluation, NLP/LLM evaluation, or Responsible AI is highly desirable
- Certifications (Desirable): ISTQB Foundation and ISTQB Certified Tester AI Testing (CT-AI); DeepLearning.AI or equivalent ML foundations; a cloud (AWS/Azure/GCP) or MLOps certification
- Non-traditional backgrounds with demonstrable model-evaluation or ML-quality work, open-source evaluation or data-quality contributions, or competitive benchmark/Kaggle achievements will be considered in lieu of formal qualification

Experience

- 5+ years in ML or software quality assurance, test automation, data or ML engineering, or applied AI; minimum 3 years specifically in AI/ML model evaluation, LLM/Gen-AI evaluation, or ML quality assurance
- Demonstrable hands-on experience building evaluation harnesses, benchmark/golden datasets and automated evaluation suites for ML or LLM systems
- Demonstrable experience evaluating LLM/RAG systems (hallucination, groundedness, retrieval quality) and/or agentic systems (tool-use correctness, trajectory, task completion)
- Prior experience embedding evaluations into CI/CD pipelines and versioning datasets, prompts and experiments
- Prior experience delivering AI quality or evaluation work in government, BFSI, healthcare, or another regulated sector is a strong plus
- Prior experience directing, mentoring, or setting QA/evaluation standards across multiple teams is desirable, given the programme-lead nature of the role

Key Responsibilities Evaluation Strategy, Standards & Quality Gates

- Own the programme’s AI evaluation methodology and make evaluation a formal gate in the delivery lifecycle — defining, publishing and enforcing common metrics, acceptance thresholds and quality gates
- Select the appropriate evaluation approach per use case — offline benchmarks, human evaluation, LLM-as-judge, or online/production evaluation — and document the rationale
- Ensure Responsible AI evaluation is built into the standard delivery flow rather than added afterwards, aligned to the MeitY Responsible AI advisory and the IndiaAI Safe & Trusted AI framework

Evaluation Infrastructure & Reusable Harnesses

- Design and build automated evaluation harnesses, benchmark and golden datasets, and regression suites that pods reuse across ministry deployments — mirroring the programme’s build-once, reuse-many model
- Wire evaluations into CI/CD as release gates; version datasets, prompts and evaluation configurations so results are reproducible and audit-ready
- Publish reusable evaluation sets, rubrics, harnesses and templates to AIKosh under standard metadata (and to OpenForge where code is shared) for national reuse

Model & Output Quality Evaluation

- Evaluate models with task-appropriate quantitative metrics — precision/recall/F1, ROC-AUC/PR-AUC, calibration, MAE/RMSE and task-specific measures — reported with statistical rigour, including uncertainty, adequate sample sizes and vigilance against metric gaming
- Evaluate AI outputs against defined quality dimensions — accuracy, groundedness, reliability, consistency, relevance and adherence to business and policy requirements — rather than pass/fail alone

Gen-AI,



LLM & RAG Evaluation

- Assess LLM and RAG systems for hallucination, factual consistency, groundedness and citation faithfulness, relevance, prompt robustness, input validation and non-deterministic behaviour
- Measure retrieval quality and generation quality separately in RAG pipelines using RAGAS, DeepEval, TruLens or LangSmith
- Use LLM-as-judge where appropriate, calibrated against human labels, controlling for known judge biases (position, verbosity, self-preference), and defaulting to human evaluation where automated judgement is unreliable

Agentic AI Evaluation

- Evaluate agents on their trajectories, not only final answers — tool selection and tool-call correctness, multi-step reasoning and workflow completion, agent handoffs, failure recovery, memory behaviour and end-to-end task-completion accuracy
- Validate human-in-the-loop controls — review, escalation, override, feedback and exception handling — wherever AI decisions require human supervision

Safety, Adversarial & Responsible AI Testing

- Under the technical guidance of the AI Safety Researcher, operationalise adversarial and red-team findings — jailbreaks, prompt injection, indirect prompt injection, unsafe behaviour and sensitive-data (PII) leakage — into standard, reusable test suites the pods apply, and verify guardrail effectiveness (Llama Guard, NeMo Guardrails, Guardrails AI or equivalent)
- Evaluate fairness by measuring performance across demographic, linguistic, geographic and other relevant cohorts using appropriate fairness metrics (Fairlearn, AI Fairness 360)
- Author the QA-side safety and quality evidence — model cards, dataset sheets, evaluation logs and Responsible AI records — for approval

Data & Dataset Quality Validation

- Validate training, evaluation and inference data for missing values, duplicates, schema conformance, class imbalance, data leakage, distribution/covariate shift, feature consistency and train–serve skew, using tools such as Great Expectations, Evidently or Deepchecks
- Curate high-quality evaluation datasets that reflect real citizen usage, edge cases and Indian-language and regional diversity

Production Monitoring & Quality Operations

- Monitor live AI applications for quality and retrieval degradation, drift (data, concept and model), accuracy and latency regressions, token/cost behaviour, failures and safety/compliance violations
- Define alerts, thresholds and dashboards for model-quality, latency, error-rate, drift and compliance signals; support quality-incident response and feed production findings back into the evaluation sets

Human Evaluation Operations

- Design evaluation rubrics and run structured human evaluation and annotation; manage annotator quality and inter-annotator agreement; build preference/label datasets that ground and validate the automated metrics

Cross-Pod Leadership, Mentoring & Governance

- Provide technical direction to the AI QA Engineers in the team; set shared standards and harnesses,



keep pod-level practice consistent and current, and drive adoption of sound, current evaluation techniques while screening out hype
- Turn quality evidence into decision-ready go/no-go recommendations for deployments, and represent AI quality in programme governance and monthly reviews
- Contribute AI quality and evaluation requirements to NeGD RFQs and RDRs, and review agency deliverables for evaluation and quality posture

Technical Competencies

- Programming & Data: Python (advanced) and SQL; pandas, NumPy, scikit-learn; pytest for test authoring
- Evaluation & Measurement: offline benchmarks, human evaluation, LLM-as-judge and online/production evaluation; statistical rigour — significance testing, bootstrap confidence intervals, calibration (ECE) and inter-annotator agreement (Cohen’s / Fleiss’ kappa)
- LLM, RAG & Agentic Evaluation: RAGAS, DeepEval, TruLens, LangSmith or Promptfoo; separate retrieval- and generation-quality measurement; trajectory and tool-use evaluation for agents
- ML Metrics: precision/recall/F1, ROC-AUC and PR-AUC, calibration, MAE/RMSE and task-specific metrics
- Data & Dataset Quality: Great Expectations, Evidently, Deepchecks or TFDV; detection of leakage, drift and covariate shift, class imbalance, schema violations and train–serve skew
- Safety & Responsible AI Testing: guardrail testing (Llama Guard, NeMo Guardrails, Guardrails AI), PII-leakage and unsafe-output testing; fairness metrics (Fairlearn, AI Fairness 360); model cards and dataset sheets — applied under the AI Safety Researcher’s guidance
- MLOps & Observability: MLflow or Weights & Biases for experiment and dataset versioning; Prometheus, Grafana, OpenTelemetry; LLM observability (Langfuse, Arize Phoenix or equivalent); CI/CD (Jenkins, GitHub Actions)
- Test Automation (supporting): pytest as the primary automation stack; Postman or REST-assured for API testing; Selenium or Cypress where UI validation is relevant
- Security (AI-aware): PII/sensitive-data exposure, access controls around AI services, model/API misuse and data leakage, in addition to conventional application-security fundamentals (authentication/authorisation, encryption, secure APIs)
- Cloud & Platforms: AWS, Azure, GCP and IndiaAI Compute evaluation and testing environments; containerised evaluation (Docker, Kubernetes)
- Government Frameworks & Standards: IndiaAI Safe & Trusted AI framework, MeitY Responsible AI advisory, MeitY security guidelines, CERT-In directions, DPDPA 2023
- Communication & Leadership: ability to author explicit, decision-ready evaluation and quality reports for engineering, architecture and executive audiences; to set standards and mentor pod QA engineers; and to represent AI quality in cross-ministerial reviews

Desirable Evidence at Screening

- A built evaluation harness, automated evaluation suite, or benchmark/golden dataset for an ML or LLM system, with a verifiable repository, notebook or demonstrable artefact
- Documented model or LLM/RAG evaluation with quantitative results against a production or production-like system
- Open-source contribution to an ML/LLM evaluation, data-quality, or AI-observability project with verifiable commit history
- Prior AI evaluation or quality-assurance engagement for a government, BFSI, healthcare or regulated-sector client, verifiable via reference or letter of confirmation

📌 QA Lead- AI Evaluation & Quality (New Delhi)
🏢 National e Governance Division
📍 New Delhi