Selected work
Projects
Research, modeling, and machine-learning work spanning embodied AI, computer vision, language models, optimization, and predictive systems. Select any project to read the complete motivation, method, results, and contribution.
May 2025 – Oct 2025
ICRA 2026 · Accepted
LISN: Language-Instructed Social Navigation
- Introduced a benchmark for instruction-following social navigation.
- Combined VLM reasoning with real-time costmap and controller modulation.
- Achieved 91.3% average success across challenging human–robot scenarios.
Nov 2024 – Mar 2025
ICCV 2025 · Highlight
A Linear N-Point Solver for Asynchronous Tracks
- Unified structure and linear-motion estimation across arbitrary timestamps and views.
- Supported global-shutter, rolling-shutter, event-camera, and mixed sensor tracks.
- Delivered 63 μs CPU runtime with 10–30% lower velocity error.
Mar 2025 – May 2025
UC Berkeley · EECS 182
Beyond the LRH: Concept Subspaces in LLMs
- Identified limits of single-direction representations for complex semantics.
- Designed a multidimensional concept-subspace steering method.
- Improved concept alignment from 0.5/5 to 5.0/5 using five latent directions.
Feb 2025 – May 2025
MCM 2025 · Outstanding Winner & AMS Award
Rebalancing Nature’s Scale
- Built a coupled ODE model for tourism, environmental pressure, and local revenue.
- Optimized interventions and tax allocation through simulation and AHP.
- Placed Top 3 among 21,054 teams in the Mathematical Contest in Modeling.
Feb 2025
Bay Area Data Science Summit · First Place
MLP for Optimal Equity Derivative Hedging
- Led an international team in a UC Berkeley competition endorsed by Wells Fargo.
- Built an MLP-based, exposure-aware equity derivative hedging pipeline.
- Reduced test-set premium cost by 64.5% while maintaining a $10M exposure target.
Project details → · Slides · Paper · Code
Feb 2024 – Jun 2024
Machine Learning Project
Hybrid ML for Quality Prediction
- Used PCA to compress correlated physicochemical features and isolate stable signal.
- Applied Huber loss to reduce sensitivity to noisy labels and heavy-tailed residuals.
- Combined robust preprocessing with classical and ensemble predictive models.
Project details → · Slides · Paper · Code