Blog
Series A code-level walkthrough of the Parser-Free Appearance Flow Network: Feature Pyramid encoding, CUDA-accelerated correlation kernels, optical flow warping, and the ResUnet generator that composites garments onto people.
Series What the GAN-based virtual try-on system achieved, where it failed (and why), the smartphone app design, and how diffusion models are changing everything for the next generation of META FIT.
Series How OpenPose skeletal detection, Graphonomy human parsing, and custom body measurement algorithms work together to enable accurate virtual fitting — plus an exploration of PiFu for 2D-to-3D reconstruction.
Series How I tackled the universal 'what's for dinner' problem over a decade ago using classical data science — cleansing 20,000 recipes, 200,000 ingredient records, and nutritional data into a unified ML-ready dataset.
Series Using cosine similarity on nutritional vectors to find recipes that match a target meal's nutrition profile but offer completely different flavors — at both the recipe and menu level.
Series Years after building the original ML pipeline, LLMs changed everything — using the ChatGPT API to simplify elaborate recipes into weeknight-friendly meals, and reflecting on a decade-long journey from cosine similarity to LSTM to GPT.
Series Reframing meal planning as a text generation problem — using LSTM neural networks with temperature-controlled sampling to predict diverse, non-repetitive menus from historical meal sequences.
Series How an experiment where ChatGPT outperformed professional fund managers inspired me to build Senrigan, an AI stock price prediction service powered by a fine-tuned LLM.
Series My first attempt at fine-tuning ELYZA Llama-3-JP-8B on a local RTX 3060 GPU, and the hard lessons learned about VRAM limitations when training large language models.