Vision backbone
Keural Vision Encoder · Mid
Adaptive token budgets, hierarchical concept tokens, and continuous 2D positional encoding. Checkpoint weights and saved evaluation files are available.
Model & checkpoints ↗Implementation ↗Vision · Multimodal learning · LLM systems
Models, methods,
and evidence to inspect.
My work connects biomedical AI research with adaptive vision encoders, Korean–English language-model workflows, and reproducible engineering tools.
Released model artifactsPublic source codeRecorded evaluationsThis page explores published artifacts. It does not run model inference in your browser.
Flagship model work
Bilingual multimodal system
Keural VLM · Mid
A frozen vision encoder, level-aware projector, and Qwen decoder with LoRA adapters. The release contains projector and adapter weights, not the full base decoder.
Projector & adapters ↗Inference & evaluation ↗Language-model deployment
Keural Cortex · 8B
General-purpose Korean–English MKD release with model weights, a chat template, and deployment documentation. Original MKD attribution is preserved.
Model card & weights ↗General-purpose model; not clinically validated.
Actual saved artifacts · PoC encoder
Saliency & token placement
Select a recorded visualization exported by the PoC vision-encoder project. These are saved outputs, not newly generated predictions.
PoC checkpoint: step_36k. This is distinct from the Mid encoder above.
Inspect original file ↗Bilingual VLM design
Frozen backbone,
trainable alignment.
Technical architecture ↗Published evaluation JSON · Mid encoder
Results tied to a checkpoint.
Displayed values are from checkpoint-18000 with token_budget=512. They are not independent reruns, cross-domain guarantees, or clinical results.
—
ImageNet Top-1
—
Flickr30K image→text R@10
—
CIFAR-100 linear probe Top-1
Show pinned revision and evaluation data
The model card reports limitations and comparisons with stronger pretrained baselines. Inspect the evaluation method and intended workload before reuse.
Engineering you can explore
Dataset tooling
LLM Dataset Inspector
Inspect structure, duplicates, formatting, and split leakage with evidence and remediation suggestions.
Open hosted audit app ↗Source & examples ↗Existing company-account app; may need to wake. Deployment was not modified for this showcase.
Training infrastructure
MKD Fine-Tuning Platform
Gradio UI, FastAPI workflow endpoints, GPU fine-tuning, evaluation, and checkpoint export.
Architecture & setup ↗GPU training is not hosted here. No new training run or performance result is claimed.
Research foundation
Biomedical AI and multimodal learning
My publication portfolio spans wearable PPG/EEG sensing, transformer-based stress classification, and medical-image analysis. Selected work appears in IEEE JBHI, IEEE Sensors Journal, and IEEE BioCAS.
Publications and research profile ↗