student / researcher / builder

Supervised Program for AI Research

• Building a safety benchmark for LLM/agent shutdownability in realistic tool-use tasks: scenario suite, instruction hierarchy, and failure-mode taxonomy.
• Implementing automated red-team agent harnesses with provider SDKs plus trace scoring to detect shutdown resistance, including delay, deflect, evasion, and goal-preservation behavior.

Walled AI

• First author, EMNLP Industry Track 2025. Built FINRISKEVAL, a dataset and eval pipeline spanning 1,720 profiles, 8 models, and 13k+ outputs to measure correctness and intent alignment in finance, and ran 100+ prompt and scoring ablations.
• Co-authored IMDA-commissioned LLM deployment playbooks covering red teaming, domain safety evals, intent alignment, and guardrails for production systems.

University of Bristol (UoB)

• Built eval harnesses for latent alignment failures via representation backdoor attacks, fine-tuned attacker and defender models, tested mitigations, and quantified reliability under white-box and black-box conditions.
• Threat-modeled 5+ attack vectors and measured sleeper-agent persistence and mitigation breakpoints across settings.

Emsec Private Limited

• Built and operated a honeypot fleet simulating 7k+ vulnerable applications to capture real-world attacker reconnaissance.
• Investigated live attacker activity across several honeypots, curating datasets and labeling adversarial activity to strengthen risk-aware defensive automation.
• Implemented a telemetry-to-evidence pipeline to parse, enrich, and aggregate attacker logs into usable detection signals.

Emsec Private Limited

• Optimized C++ DSP modules for high-throughput SDR pipelines and integrated real-time ingestion and processing for low-latency inference, with profiling and latency tuning.