cv

Pre-doctoral Researcher

Google DeepMind

Conducted research on active foundation model evaluation, mechanistic interpretability in VLMs, and cultural alignment for Gemini.

• ProEval: Developed an active evaluation framework that estimates model performance and discovers failures using Bayesian inference and Gaussian Process surrogates, reducing evaluation cost by 65x while uncovering 5x more diverse failures across 16 LLMs/VLMs (Advisor: Dr. Zi Wang).
• Steering for Multilingual Fact Recall in VLMs: Conducted activation patching to identify cross-lingual failure circuits, finding late-layer attention routing causes multilingual gaps, and showed text-derived steering vectors recover up to 17% multilingual accuracy (Advisor: Dr. Laura Rimell).
• Improving Cultural Awareness in Gemini: Improved cultural alignment across 6 locales via reference-guided critique reward modeling for RLHF, and designed an adaptive agentic information-theoretic evaluation framework (Advisors: Shachi Dave, Dr. Partha Talukdar, Dr. Bidisha Samanta).

Research Assistant

Social AI Studio, SUTD

Researched low-resource NLP safety benchmarks and multimodal meme analytics under Dr. Roy Ka-Wei Lee.

• SEAHateCheck: Built the first functional hate speech detection benchmark for 8 Southeast Asian languages across 19 functional tests, engineered a LoRA PEFT benchmarking pipeline for 12 LLMs, and quantified sociolinguistic gaps on protected categories.
• MATK: Built a unified hateful meme classification toolkit across 8 datasets and 7 models with YAML experiment orchestration, and used Integrated Gradients to diagnose VLM biases.

Software Engineer Intern

Publicis Sapient

Built conversational AI systems and cloud observability pipelines.

• Built a RAG conversational AI system with LangChain and Kafka pipeline for real-time city data ingestion into a vector store.
• Implemented RESTful APIs for multi-turn conversation memory, concurrency, and context.
• Set up Prometheus-Grafana observability stack and GitLab CI/CD pipeline reducing build lead time by 2 hours.

Full Stack Developer Intern

SAP AI Labs

Developed document processing microservices and security automation workflows.

• Expanded document processing capabilities by integrating barcode and QR code recognition across 8 document formats.
• Redesigned microservices with domain-driven design to improve scalability and reliability across 3 services.
• Built automated Docker vulnerability triaging workflows using Protecode and Groovy.

Master of Science in Intelligent Information Systems (MIIS)

Carnegie Mellon University

School of Computer Science

Bachelor of Engineering (Honours with Distinction)

Singapore University of Technology and Design (SUTD)

Computer Science; Minor in Artificial Intelligence

• Semester Abroad at University of Maryland, College Park (Semester Academic Honours)

Steering for Multilingual Fact Recall in Vision-Language Models

AAAI '27 (Under Review)

Anonymous Authors (1st Author - A. Kumaresan)

ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation

ICML '26

Y. Huang, W. Zeng, A. Kumaresan, Z. Wang

SEAHateCheck: Functional Tests for Detecting Hate Speech in Low-Resource Languages of Southeast Asia

ACM TALLIP '26

R.C. Ng*, A. Kumaresan*, Y. Hu, R.K. Lee (* Equal contribution)

Bridging Modalities: Enhancing Cross-Modality Hate Speech Detection with Few-Shot In-Context Learning

EMNLP '24

M.S. Hee*, A. Kumaresan*, R.K. Lee (* Equal contribution)

MATK: The Meme Analytical Toolkit

ACM MM '23

M.S. Hee, A. Kumaresan, N.K. Hoang, N. Prakash, R. Cao, R.K. Lee

Google Pre-doctoral Research Program

Google DeepMind

Selected as 1 of 12 researchers from over 30,000 applicants for the 2025 cohort.

Google DeepMind Recognition

Google DeepMind

Awarded 3 peer bonuses and 1 spot bonus for research contributions across different projects.

3rd Place, Cisco Webex Hackathon

Cisco

Won 3rd place out of 250 teams for the project Hidden Emergency Contact System.