Hi, I'm Laksh.
I work across machine learning, research engineering, and scientific computing. Recently, I’ve been interested in agentic LLMs, computational neuroscience, quantitative modeling, and applied scientific ML.
At Hillclimb (YC F25), I trained mathematical reasoning in Hermes 3 and Hermes Agent systems.
With Berkeley AI Research Labs / Dharmamitra, I worked on OCR and text-normalization pipelines for Sanskrit and Tibetan translation systems.
Previously, as a Velexi Research Scholar at Velexi Research, I worked on scientific ML for IR spectra, including signal processing, functional group classification, and dictionary-learning style approaches.
Highlighted Papers:
The Geometry of Forgetting: CATS@ICML 2026 [Oral]
Predicted actual alignment degradation with R² = 0.991; diagonal Fisher approximated full Fisher with R² = 0.887.
Symmetry-Constrained Gaussian Processes: ICML 2026 AI for Science [Oral]
Matched baseline molecular-property accuracy with up to 5x fewer labels; reduced FreeSolv MAE by 31.7% at 1600 labels.
ShapeUQ: CVPR 2026 Workshop 3D4S [Oral]
Produced 90% confidence intervals for PDE simulation error while running 14x - 31x faster than Monte Carlo.
Adaptive Meta-Curriculum for Test-Time Self-Improvement: ICLR 2026 Workshop RSI [Spotlight]
Improved test-time compute efficiency by 2.3x and raised math reasoning accuracy by 18.7%.
Data Cartography for Detecting Memorization Hotspots: ICML 2025 DIG-BUG Workshop [Best Poster]
Reduced canary extraction by >40% with only 10% pruning and <0.5% validation perplexity increase.
For further inquiries: lpatel [at] caltech [dot] edu
Tech Stack:
