Projects
Selected work with links.
BigTB6 — multimodal TB screening
2026Medical AI — Multimodal architecture fusing self-supervised HeAR acoustic embeddings with MedGemma + MedSigLIP vision-language models, with domain-aware expert routing for natural vs. forced coughs. 0.75 patient-level AUC, 5-fold CV across 33K samples. 2nd place, Agentic Track — MedGemma Impact Challenge ($5,000, 850+ teams).
HeAR · MedGemma · MedSigLIP · PyTorch
ORCA-G4 — on-device oral cancer screening
2025Medical Vision AI — YOLO11 lesion-detection pipeline (273 images → 362 crops, up to 0.553 mAP@50), then Gemma fine-tuned with LoRA + 4-bit quantization on a single RTX 3090 for offline vision-language classification with JAX-exported on-device deployment.
YOLO11 · Gemma · LoRA · JAX · 4-bit
IndicPHI-Bench — multilingual eval framework
2025Evaluation — Reproducible evaluation framework spanning 23 Indian languages and 13 scripts; fine-tuned a 237M-parameter MuRIL encoder for 50 entity types to study multilingual transfer and character-level leakage.
MuRIL · Multilingual NER · Eval Harness
GPU kernel optimization / inference engineering
2025Systems — Progressively-tiled CUDA GEMM kernels (memory hierarchy, register tiling, occupancy); extending to attention and inference optimization with a focus on CUDA and FlashAttention.
CUDA · GEMM · FlashAttention
More → github.com/SACHokstack