Changelog & Milestones

Fine-tuned Qwen2-VL-2B (QLoRA) for explainable diabetic retinopathy grading on a 6 GB consumer GPU, 66% accuracy, 0.76 QWK. Published the model and a 2,064-pair clinical VQA dataset on Hugging Face, with an ACL-format paper and offline Gradio demo.

Built a fully reproducible fake-news classifier (TF-IDF + Passive-Aggressive) reaching 0.99 F1, with unit tests, GitHub Actions CI, a Gradio demo, and a candid write-up on source leakage and model limitations.

Joined the Vicharanashala Lab for Education Design under Dr. Sudarshan Iyengar, deploying AI/ML pipelines targeting gaps in India's education ecosystem with direct commit authorship on live open-source repositories.

Recognized among the top 2% out of 40,000+ participants nationwide (GUVI-HCL) for innovation, technical excellence, and AI-driven problem-solving.

Completed a BERT/SBERT-based grading pipeline with OCR for handwritten answers, cutting manual evaluation time by over 70%. Built under Dr. Ashwini Kumar Upadhyay, REC Kannauj.

Developing a hybrid architecture for low-resource language modeling, balancing memory footprint and inference latency without sacrificing accuracy.

Led technical outreach for Google Gemini across campus, organizing AI workshops and tool integration sessions.