Projects

I learn sparse combinations of pretrained channels for a target task, then fold the mixing coefficients into ordinary model weights. Experiments cover vision models and VLM image encoders.

I measured adaptation time on an NVIDIA RTX PRO 6000 GPU.

I use Python, PyTorch, torchvision, and Hugging Face Transformers to search layer capacity, pruning operators, and recovery settings. CUDA runs the GPU experiments.

I compare throughput, latency, and memory on RTX PRO 6000 GPUs and an AMD EPYC CPU. The output is a compact dense encoder.

I use meta-learning to retrieve adapter placement, rank, and initialization for a new medical-imaging dataset.

I used PyTorch and MONAI for medical-image experiments and supernet training. Dataset and subnetwork embeddings let a new task retrieve an architecture and its initial weights.

I used ONNX Runtime and TensorRT to run railway anomaly detection on edge devices, with profiling to check inference cost.

I used PyTorch to implement diffusion-based architecture search over network configurations.

I used PyTorch and multi-objective Bayesian optimization to search hybrid CNN–Transformer architectures. I profiled models on edge devices and worked with INT8 quantization.

I studied how to search for efficient Vision Transformers for object recognition on small edge devices.

I used PyTorch for mixture-of-experts architecture search, balancing accuracy and latency while considering fairness and robustness.

I used JAX for complex-valued physics-informed networks that model quantum spin systems with physical constraints.

I’m exploring a hypernetwork that uses prompts and budget constraints to generate structured attention and FFN masks for vision encoders.

I built distributed training pipelines on NVIDIA and AMD ROCm clusters, using FP16/BF16 precision and multi-GPU profiling for large vision models.

I used Python, Streamlit, and MLflow to compare model metrics and experiment runs in a dashboard.

I implemented flowshop scheduling methods in C++: branch and bound, CDS, NEH, Chen, simulated annealing, tabu search, and genetic search.

Qt lets users choose algorithms and benchmarks and compare makespan and runtime. OpenMP parallelizes search; Python notebooks support the experiments.

I used OpenMP parallel loops, private variables, and reductions in C, with thread counts and execution timing for numerical integration.

I defined request and response types in an RPC interface, used rpcgen for the stubs and XDR serialization, and wrote C client/server code with error handling.