Data Science student building systems that solve real problems with data, NLP, and deep learning. Focused on production-ready projects with live demos and measurable impact.

Python Pandas NumPy TensorFlow PyTorch SQL Git

I am a final-year DataScience student at University of Engineering and Technology Peshawar, specializing in data science and natural language processing. My work bridges traditional data analytics with modern AI — from Power BI dashboards for business insights to RAG-based systems for sensitive domains like religious text retrieval.

I build projects end-to-end: data collection, cleaning, modeling, deployment, and documentation. Every repository includes a live demo, professional README, and reproducible code.

A production-deployed Retrieval-Augmented Generation system that answers Islamic jurisprudence questions using authenticated Quran, Hadith, and Tafseer sources. Constrains LLM output to retrieved content only — eliminating hallucination risks critical for religious guidance.

• Architecture: multilingual-e5-large embeddings + ChromaDB vector search + Llama 3.3 70B constrained generation
• Corpus: 36,606 documents (6,236 Quran verses, ~30,000 Hadith, 6,235 Tafseer entries)
• Performance: 3-4s response time, 100% source citation accuracy, 0% hallucination rate by design
• Live Demo: huggingface.co/spaces/noormrc123/fiqah-qa
• Repository: github.com/Mahnoor-data/QuranFiqah

Key technical decisions:

• Selected ChromaDB over Pinecone/Weaviate for zero-cost persistent storage
• Implemented batch embedding (64 docs/batch) to optimize Colab GPU utilization
• Designed strict system prompts with temperature=0.3 to prevent creative generation of religious rulings
• Handled real-world data engineering challenges: API failures, PDF OCR limitations, duplicate ID resolution

Deep learning comparison of RNN, LSTM, GRU, BiLSTM, and DistilBERT for sentiment classification of code-mixed Hinglish (Hindi-English) YouTube comments. Includes full data pipeline from collection to visualization.

• Dataset: 24,000 comments from 12 videos, balanced to 9,552 samples
• Models: RNN (65%), LSTM (72.1%), GRU (72.5%), BiLSTM (74%), DistilBERT (85%)
• Key Finding: BiLSTM delivers best accuracy-time ratio (74% in 9.9s); DistilBERT dominates accuracy but costs 28x training time
• Techniques: Automated RoBERTa labeling, stratified split, custom readability filters, professional matplotlib visualizations
• Repository: github.com/Mahnoor-data/hinglish-sentiment

End-to-end retail analytics project analyzing consumer behavior, outlet performance, and product trends.

• Tools: Python (Pandas, NumPy), SQL, Power BI
• Impact: Identified Low Fat products = 64% of sales; Medium outlets = 44% revenue engine
• Deliverable: Interactive Power BI dashboard with drill-down capability
• Repository: github.com/Mahnoor-data/E-Commerce-Sales-and-Insights

Unified 7 relational tables into a customer model to identify churn drivers and retention opportunities.

• Tools: Power BI (DAX), SQL
• Impact: Discovered 104K churned customers causing 9.38% revenue loss; 147K at-risk customers identified
• Key Drivers: Order cancellations and late deliveries
• Deliverable: Power BI dashboard with churn risk segmentation
• Repository: github.com/Mahnoor-data/Customer-Retention-and-Churn-Analysis

Statistical evaluation of campaign effectiveness across channels and devices.

• Tools: Python (Pandas, Statsmodels), Power BI
• Methods: ANOVA, T-tests, CPA/ROAS/Profit Margin engineering
• Finding: Mobile CTR > Desktop, but Desktop CVR and profitability stronger; Promos delivered highest ROAS (474%)
• Deliverable: Power BI dashboard for channel/device performance optimization
• Repository: github.com/Mahnoor-data/Marketing-Campaign-ROI-A-B-Testing

BS Datascience — University of Engineering and Technology Peshawar, Pakistan

• Email: noormrc123@gmail.com
• LinkedIn: linkedin.com/in/mahnoor-zakir-9a6183358
• GitHub: github.com/Mahnoor-data

Last updated: June 2026