satyam dev — Lead ML Engineer

I build production ML systems. My visual search and recommendation engine is live at perniaspopupshop.com Fine-tuned SigLIP2-SO400M on A100 GPUs with Multi-modal recommendation combining image embeddings + text descriptions. Users upload photos, get product matches and Similar Products Carousel - Every Product Page Shows visually similar items ranked by similarity score. Built end-to-end: training pipeline → FastAPI deployment on 2x L4 GPUs → vector database integration. Complete from model training to production API. **Technical Implementation** Model Training & Experimentation: • Tested CLIP (ViT-B/32, ViT-L/14), SigLIP v1, SigLIP2-SO400M across multiple 1-2 day A100 training runs • Custom loss: contrastive + 0.5 × image-to-image similarity • Mixed precision (FP16), gradient accumulation (8 steps), cosine annealing with warmup • Trained on 100K+ fashion images with RandomResizedCrop, AutoAugment, color jitter Production Architecture: • FastAPI with 4 workers, automatic GPU assignment (worker_pid % num_gpus) • Vector DB storing 1152-dim embeddings for 150K+ products Color Extraction Pipeline: • SCHP (Self-Correction Human Parsing) for garment segmentation • K-means clustering in RGB/LAB color space OCR System: • PaddleOCR for barcode/product ID detection Object Detection: • YOLO for multi-item images **ML Stack:** Python, PyTorch, SigLIP2/CLIP, YOLO, PaddleOCR, SCHP, Qdrant, FastAPI **Beyond ML** Lead 5 React/Vue js engineers—code reviews, architecture, feature integration.

Stackforce AI infers this person is a Machine Learning Engineer with strong frontend capabilities in the E-commerce sector.

Location: Mumbai, Maharashtra, India

Experience: 11 yrs 10 mos