Trustworthy AI for Urdu and Pakistani public information.

Peer-reviewed publications

None yet

Core papers planned

3

Public GitHub repositories

29

GitHub contributions, past year

719

GitHub figures were read from my profile in October 2026. Everything on this page is finished, in progress, or marked Planned.

Status labels used on this page: Active In progress Planned

• Urdu-RomanX

  Planned

  Do models understand Urdu script, Roman Urdu and Urdu-English text equally well?

• UrduQA-Reason

  Planned

  Can AI answer with evidence and multi-step reasoning, and say so when no answer exists?

• UrduCodeSwitch-Bench

  Planned

  How reliable and safe is AI across different forms of the same language?

• UrduIE-KG

  Planned

  Can we extract people, places, laws and relations from Urdu text?

• PAKGOV-RAG-X

  Planned

  Can AI answer Pakistani government questions with real, checkable sources? This is the planned research upgrade of the current PAKGOV-RAG prototype.

Wider interests: Nastaliq text processing, bilingual embeddings, code-switching, and low-resource fine-tuning with models such as mBERT, XLM-RoBERTa and mT5. Stretch work on RAG reliability benchmarks follows only after the core projects.

PAKGOV-RAG

In progress Version 0 pilot

A bilingual Urdu-English retrieval-augmented generation system for Pakistani government and legal documents.

The problem

Official information such as tax circulars, education policy, identity procedures, company filings and court judgments is spread across many documents and is hard for Urdu readers to search. Answers from general AI tools may sound confident without pointing to the source document.

Why it matters

People should be able to ask in the language they actually write, and check the answer against an official source.

Document areas targeted

FBR circulars, HEC policy, NADRA procedures, SECP filings and Supreme Court judgments. Source links and document IDs are kept with each document.

Approach

The repository describes four components, shown below in the order a document passes through them.

1. OCRTurn scanned documents into text.
2. Hybrid retrievalCombine dense and keyword search to find passages.
3. Citation groundingTie each answer to the passages it used.
4. EvaluationMeasure retrieval and answer quality.

Components are taken from the repository description. Detailed architecture notes will be added as they are written up.

Technologies in use now

Python, Git, FAISS and LangChain. I am still learning PyTorch and Hugging Face, so I do not list them as part of the working system.

Planned evaluation

Recall@5, Recall@10, MRR, nDCG, answer correctness, groundedness, citation precision and recall, unsupported claims, hallucination, abstention, speed and cost.

Measured results

None published yet. I will add numbers here only after the evaluation has actually been run, with the method and dataset described.

Known limitations

This is a pilot, not a production service. It gives no legal advice and I make no claim of legal accuracy.

Next steps

Upgrade step by step into PAKGOV-RAG-X: retrieval metrics, source grounding, an evaluation set, and a written technical report.

Open the repository

Other repositories

• python-learning-log

  Daily Python practice, programming fundamentals, algorithms and projects, documenting my progress from beginner to proficient.

  • Python

• MS-AI-Admission-Eligibility-Engine

  A reproducible, source-grounded engine that models program-specific admission requirements from official university sources and validates applicant inputs.

  • Python

• urdu-en-word-analyzer

  Bilingual text analyzer for English and Roman Urdu text, with word-frequency analysis, UTF-8 support and CSV output.

  • Python

• Pakgov-search-cli

  Command-line tool for fast keyword search across Pakistani government documents, with contextual results, search history, error handling and modular file processing.

  • Python

• ml-paper-reproductions

  Reproducing landmark machine learning papers from scratch, including Attention Is All You Need and BERT.

See all repositories on GitHub

Venues are chosen only after the real contribution is clear. My research identity is on ORCID (0009-0002-2163-1728).

Practice profiles: LeetCode, HackerRank and Kaggle.

• Python automation

  Scripts that clean files, collect public data, send reports and save manual work.

  • Python
  • Pandas
  • APIs

• Websites

  Fast, responsive business and portfolio sites with clear copy and working contact forms.

  • HTML
  • CSS
  • JavaScript

• AI chatbot and document search

  Ask questions over your PDFs and documents in English or Urdu, with the source shown for each answer.

  • RAG
  • LangChain
  • FAISS

• Data cleaning and analysis

  Turn messy spreadsheets and CSV files into clean data and clear charts.

  • SQL
  • Pandas

• Urdu and English text work

  Search, classification and processing for Urdu text, including Nastaliq documents.

  • NLP

How we work

1. Brief

   Tell me what you need. I reply with questions and a fixed quote within 24 hours.

2. Build

   I share short progress updates so you see work early, not only at the end.

3. Review

   You test it and request changes. Revisions are included in the agreed scope.

4. Deliver

   Clean code, a short guide and a handover. I stay available for fixes afterwards.

Questions clients ask

Are you experienced enough?

I am a first-year BS CS student with certificates in Python, SQL, REST APIs and JavaScript, and an AI project on GitHub. I tell you early if something is outside my skills.

How fast can you start?

Usually within two to three days of agreeing the scope. Small projects can finish within a week.

How do we communicate and pay?

Email, WhatsApp or the platform you prefer. Payment goes through Upwork, Fiverr or an agreed method, with milestones for larger work.

Do you work with Urdu content?

Yes. Urdu and English text processing, Urdu search and Urdu-English document question answering are my focus.