Hamza Jadoon

Machine Learning Engineer

PackageX • Full time

Nov 2025 - Present • 6 mos

• Vision-Language Model Fine-tuning: Fine-tuned open-source VLMs (InternVL, Qwen-VL) on logistics document images for end-to-end structured JSON extraction. Constructed supervised training pipeline with image-prompt-JSON triplet datasets; leveraged H100 GPU infrastructure on Vast.ai for distributed training with LoRA/QLoRA parameter-efficient adapters. Implemented instruction-following optimization via supervised fine-tuning (SFT) with contrastive loss objectives to align model outputs with schema-constrained JSON generation • Multimodal KVP Detection: Designed and implemented token-level sequence labeling framework using LayoutLMv3 for automated key-value pair extraction from logistics documents. Conducted comparative analysis between region-based detection (YOLOv8) and vision-language transformer approaches, optimizing for both spatial accuracy and semantic understanding of document context • Agentic Email Processing System: (POC) Architected multi-agent orchestration framework with supervisor-routing mechanism for hierarchical task decomposition. Implemented intent classification layer using prompt-engineered LLM chains with tool-use capabilities, enabling dynamic agent selection for document parsing, text extraction, and structured information retrieval from unstructured email streams • Edge Deployment Pipeline (HP Warehouse): (POC) Developed end-to-end object detection workflow with YOLOv8 for real-time drawer localization in warehouse environments. Implemented post-training quantization using Hailo8 SDK with mixed-precision optimization, achieving inference acceleration while maintaining model fidelity across diverse deployment environments