Cost-Aware Automatic Prompt Optimization (APO)

Hybrid APE-OPRO prompt optimization with 18% lower API cost and 0.84 F1.

AI Evaluation Specialist — Content Quality at Viator

End-to-end LLM evaluation framework for product descriptions: layered metrics (ROUGE, BERTScore, NLI, G-Eval, SelfCheckGPT), model introspection tooling, and CI/CD integration.

Safety & Evaluation Framework for AI Agents

End-to-end safety and behavioural evaluation framework for AI agents: trajectory scoring, safety gates, adversarial red-teaming, and continuous production monitoring.

Media Optimisation — GPT-4V + Multi-Armed Bandit Hero Image Selection

Combined GPT-4 Vision shortlisting and Bayesian Multi-Armed Bandit to replace static hero image selection with adaptive, CTR-driven optimisation at product scale.

AI Customer Service Agent for Travel (System Design)

End-to-end agent architecture with GraphRAG knowledge, multi-tier memory, HITL escalation, and sub-1.5s P95 latency.

Review Summarization at Scale

Multilingual ABSA pipeline cutting LLM token usage by 82% and hallucinations to 1.8%.

Automated FAQ Extraction (Knowledge Governance System)

GraphRAG + agentic validation system cutting hallucinations to 2.1%.

Active Learning for Traveler Tips Extraction

Fine-tuned SLM with 0.84 F1 and 99.3% lower inference cost.

Enterprise AI Governance Framework (Responsible AI)

Safety gate with 100% coverage and <10ms validation latency.

Cross-Portfolio Engineering Practices

Standardized experimentation, monitoring, and reproducibility across ML systems.

Attribution Modelling

Slideshare deck on multi-touch attribution, data quality, and modeling trade-offs.