Designed and launched 7 features for Meta Business Messaging, connecting advertisers and customers via WhatsApp.

• Implemented ad creation and messaging functionalities, including multi-WhatsApp-number chat support, ad format previews, advertiser options defaulting, WhatsApp input expansion in new bid workflows, ad delivery matcher updates, and more.
• Built full-stack web and mobile solutions using React.js, Relay, and GraphQL, and backend queries and services in Hack.
• Owned product lifecycles end-to-end by driving cross-functional design, collaboration, and experimental analysis.
• Led product decisions on features reaching 300K+ total advertisers.

Developed BountyBench, a cybersecurity bug bounty benchmark for AI agents (Principal Investigator: Percy Liang).

• Introduced the benchmark featuring 25 systems from complex, real-world codebases, and 40 bug bounties covering 9 of the OWASP Top 10 Risks, to assess LLM agents’ ability to exploit and patch security vulnerabilities.
• Built an automated verification pipeline to validate LLM-generated patches that preserve invariant functionalities in code.
• Designed end-to-end LLM agents capable of real-world security tasks, optimized with prompting, memories, and reflection.

Designed and developed the Wave Dashboard:

• Developed a full-stack ticket sales management dashboard generating $120,000+ monthly revenue using Next.js and Node.js.
• Created key features such as event-editing and real-time ticket tracking, leveraging Firebase, Google OAuth, and Cloud Run for scalable data storage, retrieval, and access management.

Supervised by Prof. Yu Huang and Prof. Kevin Leach:

• Reinforcement Learning for LLMs for Code: Generating negative code samples using RL with compiler-in-the-loop. More details coming up soon.
• Few-shot Malware Classification: Designed a semi-supervised malware classifier, pioneering efficient & semantics-aware malware data augmentation through information retrieval, invariance learning, and alignment techniques. The classifier achieved SOTA results with 5.60% improvement in few-shot settings and 11.44% improvement under concept-drift.
• Merging Human & Transformer Attention: Implemented a Transformer-based NLP model whose self-attention weights are refined using human attention data captured through eye-tracking. The model is applied to neural code summarization tasks with a 29.91% performance improvement, showcasing the potential of integrating human attention into neural models.
• LLM Feature Attribution Analysis: Comprehensively evaluated feature attribution in six LLMs (GPT, Llama 2, StarCoder, etc.) in code summarization through Shapley-value analysis; compared against human eye-tracking attention patterns.

Supervised by Prof. Lingming Zhang:

• Designed a universal LLM-based fuzzer targeting compilers of any programming language, with LLM-mutated code inputs integrated in the fuzzing loop; tested on 7 languages with significantly improved coverage and uncovered 76 new bugs.
• Specialized in JIT vulnerabilities within the V8 JavaScript engine; refined fuzz testing algorithms for targeted exploitation; offered expertise in instrumentation, auto-prompting, and compiler and systems programming for vulnerability pinpointing.

Post-Investment Management Team:

• Worked on the post-investment team to eliminate future manual processing of three portfolio companies’ operational data.
• Implemented a full-stack web service in React JS, Flask, and MySQL for robust client data catalog collection and storage.
• Automated the conversion of raw operational data into financial statements using OpenPyXL, Pandas, NumPy, and Seaborn.
• Built an interactive sales performance visualization interface using D3 for real-time strategic insights.