✦ golden hour
Yuvraj Verma
AI / ML Engineer · NLP & Computer Vision Researcher
I'm an AI/ML engineer powered by curiosity and a mild obsession with making things work better. I turn messy, real-world problems into clean systems — and pick up something new on every detour. Builder by instinct, researcher by habit, designer when no one's watching. Off the keyboard I run on big ideas, good stories, and the occasional Harvey Specter monologue — because I don't have dreams, I have goals.
Projects shipped
Internships
India AI Buildathon
Building with AI
A bit about me
I'm Yuvraj — an engineer who likes turning research into things that actually run. My work lives at the intersection of machine learning, natural language processing, and computer vision, with a soft spot for AI that makes learning and safety smarter. I care about clean systems, honest results, and shipping things people actually use.
Where I've been
AI / ML Research
Education-focused open source
Building AI/ML for India's classrooms, one commit at a time.
ML & Full-Stack Intern
Reaching Sky Foundation
Taught a BERT model to read between the lines for mental-health detection.
Machine Learning Intern
BreakOut AI
Built an LLM tool that reads the web so humans don't have to.
B.Tech, Electronics Engineering
REC Kannauj
Where circuits met curiosity. CGPA 7.5.
BS in Data Science (Foundation)
IIT Madras
The math-and-models foundation under everything I build.
Machine Learning Intern
Codesoft Technologies
My first taste of real datasets, classifiers, and that EDA rabbit hole.
Things I've built
IDRiD-VQA: Explainable Retinopathy Grading
Qwen2-VL-2B + QLoRA fine-tuned on a 6 GB GPU to grade diabetic retinopathy and explain its reasoning — 31%→66% accuracy, 0.76 QWK, model & dataset on Hugging Face.
Fake News Detection Pipeline
Reproducible TF-IDF + linear-model classifier hitting 0.99 F1, with tests, CI, a Gradio demo, and an honest analysis of what the model actually learns.
Automated Answer-Sheet Evaluation
BERT/SBERT + OCR pipeline that grades descriptive answers and cuts manual evaluation time by 70%+.