Shankari Venkatesh

Contributed to the evaluation and quality assurance of Google’s Gemini platform within a cross-functional team focused on large-scale LLM reliability and safety. Strengthened Gemini’s coding benchmark programme by authoring original solutions and delivering peer-level code reviews across multiple languages, improving benchmark coverage by 25% and output reliability by 18%. Identified edge cases, hallucinations, and safety risks; implemented guardrail and validation frameworks that reduced unsafe response rates in targeted eval categories. Built Python-based evaluation pipelines and automated reporting tooling, cutting manual testing overhead by 30% and improving consistency across review cycles. Conducted systematic red-teaming and adversarial prompt testing to surface failure modes, feeding findings into model improvement cycles, reducing critical safety failures by 40%. Developed advanced prompt templates for multi-task LLM applications and collaborated on security compliance rev

Owned the end-to-end ML lifecycle across a live gaming platform, from problem scoping and data pipeline design through model deployment and production monitoring, achieving a 30% improvement in overall system reliability. Architected and shipped RAG pipelines for dynamic content delivery (dense retrieval with FAISS/Chroma, cross-encoder re-ranking, RAGAS-based evaluation) with a 28% lift in response relevance. Built a real-time web scraping news bot integrated with Hugging Face transformers for sentiment analysis, and designed personalised recommendations using XGBoost/LightGBM with deep ranking layers, boosting user retention and CTR. Engineered AI-driven fraud detection, containerised ML workflows with Docker, and deployed via FastAPI on AWS with CI/CD; tracked experiments with MLflow and Weights & Biases.