Huzefa Merchant

I'm an ML Engineer focused on Generative AI, Multimodal AI, and Computer Vision, building real systems trained on real benchmarks like Flickr30k and COCO 2017. I rebuilt the same image captioning system 7 times — from InceptionV3+LSTM to ConvNeXt+Perceiver+GPT-2 — because getting it right mattered more than getting it done.

In my flagship LensToWords multimodal captioning pipeline, I designed an end-to-end model (ConvNeXt-Tiny → Perceiver with 16 latents → cross-attention injected into all 12 GPT-2 blocks) and drove it through phased training. I achieved BLEU-4 = 0.1341 on COCO 2017 after training on Flickr30k for architecture validation, then scaling up to COCO.

I also ship deployed ML applications: a Streamlit Movie Recommender System with a production similarity matrix, a FastAPI + Streamlit Potato Disease Detector structured as a deployable two-layer service, and NLP/vision products like the WhatsApp Chat Analyser and Fashion Recommender System. Built through hands-on research and Stanford/DeepLearning.AI certifications, I iterate with diagnosis-first engineering and I'm motivated by measurable outcomes, not demos.