rouge-score

An end-to-end MLOps project for text summarization using the HuggingFace Pegasus model. Includes a full training pipeline, evaluation, and a FastAPI for deployment.

Information Retrieval Project : Crisis Facts Summarization

Develop a chatbot that can effectively adapt to context and topic shifts in a conversation, leveraging the Stanford Question Answering Dataset to provide informed and relevant responses, and thereby increasing user satisfaction and engagement.

Comparing QLoRA, Prompt & Prefix Tuning on Mistral-7B for medical instruction-following

End-to-end MLOps pipeline that catches LLM quality regressions before production. Every PR is scored against a versioned golden dataset using BERTScore + ROUGE-L + an LLM-as-Judge rubric, compared to the MLflow production baseline, and shadowed against 5% of live traffic. FastAPI + Celery + TimescaleDB + Streamlit + DVC + GitHub Actions.

Geography-aware news aggregation system that classifies news articles by geographic regions and generates region-specific summaries.

Benchmarking framework for evaluating and comparing LLM outputs using ROUGE, BLEU, and LLM-as-judge scoring with a multi-model comparison dashboard

Python implementation of a 7-metric hybrid extractive text summarization algorithm on the BBC News dataset, outperforming baseline ROUGE scores.

📰 End-to-end NLP pipeline for news intelligence — fine-tuned RoBERTa multilabel classifier, spaCy NER, T5/BART entity-aware summarization & 5-signal misinformation risk scoring. Served via Streamlit. 🧠📊🚀

English-to-Hindi Neural Machine Translation using a Seq2Seq GRU Encoder-Decoder with Additive Attention. Built from scratch with TensorFlow, teacher forcing, XAI (Integrated Gradients), and BLEU/ROUGE evaluation to improve translation quality.

Automated LLM evaluation framework for a fictional ShopMate customer support chatbot using golden datasets, keyword matching, ROUGE-L, LLM-as-Judge, prompt variants, and model comparisons.

Research-grade RAG benchmarking platform with hybrid retrieval, sentence-level grounding, hallucination analysis, and quantitative model comparison.

LLM evaluation on a 643-question Austrian tax-law benchmark: ROUGE, BLEU and BERTScore plus a manual failure-mode analysis. Compares a QLoRA fine-tuned 5B model against a 31B zero-shot baseline; the smaller fine-tuned model wins on all five metrics.