cv Download CV

Stanford University (Lall Group)

Research at the intersection of machine learning, optimization, and control.

• Developed TaperNorm, a gated normalization layer that tapers to a foldable affine map, enabling removal of per-token normalization inside Transformer blocks and faster inference after weight folding (up to 1.22× throughput on an H100 microbenchmark).
• Learned explicit MPC policies with hard closed-loop safety constraints by training neural network controllers via an augmented Lagrangian / primal-dual loop; achieved constant-time policy evaluation (~0.035 ms worst case) with zero safety violations in simulation test sets.
• Developed and analyzed a primal-dual Adam-style method for equality-constrained optimization used for constrained neural network training, clarifying how step sizes interact with constraint geometry.

Inflection AI

Owned an end-to-end research project testing whether a tokenizer-free Byte-Latent Transformer (BLT) can be adapted to new languages by retraining only the ~4% "interface" modules while keeping the latent transformer fixed.

• Fine-tuned BLT by training a new entropy model and updating the local encoder/decoder while freezing the latent transformer (and hash n-gram embeddings), training ~4% of parameters.
• Built training and evaluation pipeline (FineWeb-2) and benchmarked on Belebele; analyzed entropy-threshold patching as an explicit compute/quality knob.
• Implemented character-preserving patch boundaries (UTF-8 safe) and studied patch-count efficiency vs. accuracy.

TSMC

Injected domain expertise into deep learning models (CNNs) via invariance constraints with exact-fit requirements on critical samples.

• Developed a training method to enforce expert invariances by requiring exact fit on a designated subset of rare-but-critical samples while maintaining performance on the full distribution.
• Built custom training and evaluation tooling to measure constraint satisfaction and robustness across regimes.

KLA Corporation (FaST Division)

Inverse problem tooling and algorithm optimization for semiconductor metrology.

• Developed a MATLAB GUI prototype for the Axion tool to run a new inverse algorithm and integrate it into an existing workflow.
• Improved algorithm accuracy by 2× and reduced runtime by 3× by optimizing signal processing and regression components.

Applied Materials (AIx Team)

Fault detection and analytics for semiconductor manufacturing using process data.

• Analyzed process traces to translate qualitative failure modes into concrete definitions and measurable features.
• Engineered fault detection algorithms for an advanced data analytics product and contributed to pipeline development.

Stanford, CA

Stanford University

Electrical Engineering (Advisor: Prof. Sanjay Lall)

Research across LLMs/Transformers, constrained optimization, and control. Expected graduation: June 2026.

Stanford, CA

Stanford University

Electrical Engineering

M.S. earned en route to the Ph.D., with coursework focused on optimization, control, probabilistic modeling, and ML.

• EE364A — Convex Optimization I
• EE263 — Introduction to Linear Dynamical Systems
• CS229 — Machine Learning
• CS228 — Probabilistic Graphical Models
• EE276 — Information Theory
• AA203 — Optimal & Learning-Based Control
• EE367 — Computational Imaging

Cambridge, UK

University of Cambridge (Pembroke College)

Chemical Engineering

Graduated #1/60 in final year. Master's thesis on robust MPC for safe/efficient exothermic batch processes.

Stanford University

TA for a core linear algebra course focused on SVD, least squares, and matrix methods used throughout ML/control.

• Held regular office hours and helped students build intuition for SVD, conditioning, and least squares through worked examples.
• Delivered a number of lectures to ~80 students, including prepared slides and in-class problem solving.
• Designed and ran midterm and final exams (writing problems, coordinating logistics, and grading).

California Institute of Technology

Numerical simulations studying interactions between acoustics and laminar flames.

University of Cambridge

Identification of metal-organic frameworks with desired structures in the Cambridge Structural Database.

Preprint

Introduced TaperNorm, a gated RMSNorm/LayerNorm replacement that transitions to a foldable affine map for faster inference; analyzed stability via output scale anchoring.

IEEE Conference on Control Technology and Applications (CCTA)

Analysis and convergence results for a constrained optimization algorithm for NN training.

UKACC 14th International Conference on Control (CONTROL)

Offline constrained training to approximate explicit MPC policies while enforcing feasibility/safety constraints.

Computers & Chemical Engineering

Combined Hamilton–Jacobi reachability with MPC to avoid thermal runaway in exothermic batch processes.

Computers & Chemical Engineering

Robust MPC for improving safety and efficiency of exothermic batch process operation.

University of Cambridge

Ranked #1/60 in 4th year (Chemical Engineering Tripos).

Pembroke College, University of Cambridge

Awarded annually for First Class Honours (Years 2–4).

University of Cambridge

Top 5 ranking in 2nd year examinations.

Pembroke College, University of Cambridge

First Class Honours in 1st year examinations.

Pembroke College, University of Cambridge

Excellence in both Tripos examinations and University-level sport.

University of Cambridge

Best poster and presentation at the 4th-year student conference.