Selected research and engineering work across multimodal systems, efficient reasoning, AI for science, and control.

Multimodal Systems

Work on vision-language models, structured generation, and reliable multimodal outputs.

Vision-Language Model Fine-tuning for Schema-Constrained Food Extraction

Completed

Fine-tuned a vision-language model for deterministic food understanding as schema-constrained JSON generation.

Built a 1,500-sample instruction dataset, added schema validation, and released reproducible training assets and code.

Code
Vision-Language ModelsStructured GenerationParameter-Efficient SFTSmolVLM2-500MPEFTAdamW

Reasoning and Efficiency

Studies of recursive reasoning, adaptive computation, and compute-aware model behavior.

Efficiency Analysis of Tiny Recursive Models for Reasoning Tasks

Completed

Reproduced Tiny Recursive Models on Sudoku-Extreme and studied the compute-accuracy trade-offs of recursive reasoning.

Found an overthinking regime under ACT and identified a shallow H2-L2 design as the best fixed-budget configuration.

Recursive ReasoningAdaptive Computation Time (ACT)Efficiency EvaluationTransformers

Scientific and Control Systems

Projects at the intersection of scientific modeling, long-context inference, and dynamical control.

GeneLM-Evo2: Zero-shot Genomic Variant Pathogenicity Analysis

Completed

Built a zero-shot genomic variant scoring workflow around Evo2 for long-context pathogenicity analysis.

Deployed scalable H100 inference, integrated ClinVar-based evaluation, and shipped both code and a public project site.

Code Website
Genomic Foundation ModelsZero-shot InferenceML SystemsAI for ScienceEvo2 7BModalNVIDIA H100

Optimal Control of Interleukin-10 Immune Modulation

Completed

Modeled IL-10 immune regulation as a continuous-time control system under physiological noise.

Designed a Kalman filter plus LQR controller and evaluated closed-loop behavior in MATLAB/Simulink.

ControlState-Space ModelingFiltering & OptimizationKalman FilterLQRMATLAB/Simulink