AI Engineer focused on Computer Vision applications. Experienced in building end-to-end ML pipelines from research to production deployment.
Core Expertise: Deep Learning, Computer Vision, MLOps, Model Optimization
Languages
ML/AI
DevOps
Advanced classification system for 500-class sketch images using Vision Transformers with data-centric optimization.
| Metric | Baseline | Final | Improvement |
|---|---|---|---|
| Accuracy | 50.3% | 90.3% | +40.0pp |
| Training Speed | 1x | 4-6x | AMP Optimization |
| Memory Usage | 100% | 80% | Attention Freezing |
Tech: DeiT3 ViT PyTorch Label Smoothing TTA AMP
Pixel-level semantic segmentation for 29 bone structures in hand X-ray images.
| Metric | Result | Performance |
|---|---|---|
| Dice Score | 97.64% | Clinical-grade accuracy |
| Training Time | 50% faster | AMP implementation |
| Reliability | Consistent | Cross-patient validation |
Tech: U-Net++ SegFormer Swin-Transformer Medical AI
Environmental AI system for identifying and categorizing 10 types of recyclable waste materials.
| Achievement | Impact | Technology |
|---|---|---|
| +5% Performance | TTA on Swin Transformer | Weighted Boxes Fusion |
| Real Deployment | Pilot recycling facilities | Edge optimization |
| Data Innovation | Diffusion model augmentation | Class imbalance solution |
Tech: YOLO Swin-Transformer Diffusion Models WBF Ensemble
Pure data-centric approach for receipt text detection with fixed EAST architecture.
| Phase | F1 Score | Improvement | Innovation |
|---|---|---|---|
| Baseline | 0.20 | - | Raw dataset |
| Optimized | 0.8321 | +315% | Data-centric AI |
| Efficiency | Same quality | 50% faster | Pipeline optimization |
Tech: EAST Data-Centric AI Albumentations Pipeline Optimization
SALMONN-based multimodal large language model optimization for efficiency and performance balance.
| Optimization | Before | After | Improvement |
|---|---|---|---|
| Memory Usage | 9.18GB | 5.96GB | -35% |
| Audio Captioning | 0.20 | 0.32 | +58.8% |
| Speech Recognition | 15.2% WER | 14.0% | -7.7% error |
Tech: SALMONN Llama-3 4-bit Quantization Flash Attention 2 VB-LoRA
|
Contribution Stats
|
Language Distribution
|
Contribution Calendar
Contact
