Harshvi User

I’m a Computer Vision & AI/ML Engineer with 1+ year of production experience building deep learning pipelines from scratch. I designed and trained a custom Multi-Stage CNN with heatmap regression for 68-point facial landmark detection, achieving 98% accuracy with 0.10ms inference on 100K+ images.

I’m hands-on with PyTorch and OpenCV, and I’ve built end-to-end 2D-to-3D face reconstruction using a FaceVerse 3D morphable model, with interactive rendering via Three.js. I also developed 3D hair mesh generation using monocular depth estimation and semantic segmentation, reconstructing dense point clouds and triangular meshes with Open3D and Trimesh.

I integrate LLMs in production workflows too—building RAG pipelines (LangChain + Supabase vector DB/pgvector) for grounded, document-scoped answers, and fine-tuning TinyLLaMA and Qwen via LoRA on Hugging Face Transformers with 40% GPU compute reduction. I’ve also shipped real-time streaming systems like a voice assistant (STT → LLM → TTS) using WebSockets, cutting end-to-end latency from 890ms to 200ms.