Experience

1. Research Assistant

   Jan 2025 — Present
   Q-Lab, UCSD

   Advisors: Prof. Lianhui Qin and Matthew Ho
   • Contributed to ArcMemo by leading the design of a program-synthesis-style memory ontology, developing many-to-one puzzle-to-feature mappings and manually curating concept parameterizations to enable concept-level reasoning.
   • Engineered a reasoning-based retrieval mechanism (System-2 exploration) to resolve embedding failures, achieving a 7.5% relative gain on ARC-AGI-1 (59.33% official score).
   • Built a complete concept dataset generation pipeline transforming hand-written concepts into validated helper puzzles through multi-stage LLM-based generation, code synthesis, and automated testing.
   • Extended the framework to AIME math problems by designing a metacognitive self-assessment pipeline, improving accuracy by 9.3% via self-reflective memory usage.
   LLM Agents Program Synthesis Concept-Level Memory Python

2. Research Assistant

   Jun 2025 — Present
   Wang Lab, UCSD

   Advisors: Prof. Wei Wang and Young Su Ko
   • Developing TrustPPI: Domain-specific trust signals for protein-protein interaction prediction; showed deformation stability achieves 0.70–0.80 AUROC vs ~0.50 for generic confidence.
   • Architected a heterogeneous Mixture-of-Experts system for chemical reaction prediction on USPTO datasets (1M+ reactions), integrating four specialized expert models with learned routing mechanisms.
   • Implemented graph neural network encoders using directed message-passing architectures with shortest-path positional encodings, enabling permutation-invariant molecular representations for stereochemistry-sensitive reaction modeling.
   • Engineered training pipelines with teacher-forcing, load-balancing losses, and router warmup; designed evaluation frameworks for top-k accuracy and per-expert ablation.
   Protein Biology PyTorch Geometric MoE GNNs

3. Student Researcher

   Sep 2025 — Present
   Halicioglu Data Science Institute, UCSD

   Advisor: Prof. Hao Zhang (Senior Capstone)
   • Built an end-to-end distributed training stack for a 1.8B-parameter language model on 8 NVIDIA B200 GPUs, implementing pipeline parallelism and scaling analysis.
   • Developing TPU-optimized speculative decoding to reduce latency for test-time reasoning systems.
   Distributed Training Speculative Decoding B200 GPUs

4. Student Researcher

   Sep 2025 — Present
   Halicioglu Data Science Institute, UCSD

   Advisor: Prof. Yian Ma (Senior Capstone)
   • Developing AR-Bench, an interactive reasoning benchmark where agents decide when to ask clarifying questions versus commit to an answer.
   • Designing mutual-information-based uncertainty signals to detect epistemic uncertainty for risk-controlled stopping decisions.
   Uncertainty Quantification Mutual Information Benchmarking

5. Research Assistant

   Sep 2023 — Jun 2024
   U.S. Immigration Policy Center

   Advisors: Prof. Tom K. Wong and Dr. Gabriel De Roche
   • Built forecasting models (ARIMA, exponential smoothing) on 20+ years of naturalization data to project immigration trends for 2024 election policy briefs.
   • Developed automated data pipelines and interactive Tableau dashboards to communicate findings to non-technical stakeholders.
   R ARIMA Tableau