Applied AI · Agentic Systems · Speech AI · Computer Vision · ML Infrastructure · Developer Tooling · Data Science & Analytics

Open to conversations around AI/ML engineering, applied AI, speech systems, agentic AI, MLOps, data science and developer tooling.

I'm an AI/ML Engineer working on production-oriented artificial intelligence systems.

My work spans:

• 🧠 Machine Learning & Deep Learning
• 🤖 Agentic AI & LLM Systems
• 🎙️ Speech Recognition / Urdu ASR
• 👁️ Computer Vision
• 🛡️ AI & Developer Safety Tooling
• ⚙️ MLOps, evaluation and ML infrastructure
• 📊 Data Science & Analytics

I currently work at Punjab Information Technology Board (PITB), contributing to applied AI systems and government-scale technology projects.

Outside my organizational work, I build open-source developer tooling. My current flagship public project is OhMyDB — a fail-closed database safety proxy designed to catch risky SQL before it reaches the backend.

What interests me most is the part of AI that comes after:

model.fit()

The real engineering starts with:

data
↓
experimentation
↓
evaluation
↓
deployment
↓
monitoring
↓
human feedback
↓
continuous improvement

A fail-closed safety proxy for your database.
Catch dangerous SQL before your database has to.

OhMyDB is an open-source database safety proxy built to intercept risky SQL operations before they reach the backend while failing safely when behavior is malformed, ambiguous, or unsupported.

Python · PostgreSQL · MySQL/MariaDB · AsyncIO · SQLGlot · Docker

• Fail-closed SQL policy enforcement
• Impact estimation for risky mutations
• Prepared-statement inspection
• Transaction-state tracking and recovery
• Structural and multi-statement protection
• Audit logging and sanitization
• PostgreSQL and MySQL/MariaDB adapter architecture
• Dockerized non-root runtime

• 341 automated tests
• Python 3.11 / 3.12 / 3.13 CI
• Real PostgreSQL client/driver end-to-end validation
• Fresh-wheel installation validation
• Docker build and non-root runtime validation
• Stable wheel + source distribution artifacts
• SHA256 release checksums
• Backward-compatible legacy sql-safety-proxy CLI

v1.1.0

Primary CLI: ohmydb

👉 Explore OhMyDB · Latest Release

One of my major applied ML engineering efforts has been building and improving an Urdu ASR ecosystem around Whisper.

Whisper · PyTorch · Hugging Face · MLflow · CUDA · Linux

• Built an ASR pipeline from scratch to understand the full speech stack
• Created and expanded custom Urdu speech datasets
• Built audio preprocessing and transcription workflows
• Managed annotation and validation pipelines
• Led annotators and data-collection efforts
• Combined Common Voice, FLEURS, custom Urdu podcast data and collected live speech
• Grew the training corpus beyond 150 hours
• Fine-tuned Whisper Small, Medium and Large V3
• Built repeatable MLflow-based training and evaluation workflows
• Benchmarked models and tracked promotion-quality metrics

Whisper Large V3 → 13.61% WER

Whisper Medium → 17.91% WER

This work lives inside organizational GitLab infrastructure, so the repository is not publicly linked here.

Built and worked on a YOLO + Label Studio + MLflow human-in-the-loop pipeline for continuously improving object detection systems.

YOLO · Label Studio · MLflow · Python

• Model-assisted annotation
• Human validation loops
• Immutable annotation exports
• Dataset versioning
• Scheduled retraining
• Candidate/champion model comparison
• Model promotion gates
• Deployment-oriented CV workflows

The implementation is maintained in organizational GitLab infrastructure.

Worked on an AI system for evaluating technical reports against structured SOP / compliance requirements.

LLMs · RAG · OCR · FastAPI · LangGraph · LlamaIndex

• OCR and document parsing
• Structured SOP rule extraction
• Applicability filtering
• Evidence retrieval
• Rule-level compliance evaluation
• Source-grounded findings
• Human review workflows
• Query-letter generation
• Evaluation matrices
• Audit-ready outputs

The project is maintained in organizational GitLab infrastructure.

I’m actively building and learning deeper into modern agentic AI systems.

Areas include:

• Tool-using AI agents
• Structured LLM workflows
• Model Context Protocol (MCP)
• RAG and knowledge retrieval
• Prompt engineering
• Local LLM deployment
• AI evaluation pipelines
• Multi-step reasoning workflows
• Production-oriented agent architectures

Speech AI → Urdu ASR optimization & model evaluation
Agentic AI → Agents, MCP and tool-using systems
LLM Systems → RAG, structured workflows and evaluation
Computer Vision → Human-in-the-loop model improvement
AI Security → Safer ML / LLM / developer workflows
MLOps → Reproducible training and promotion pipelines
Developer Tooling → OhMyDB & safer database operations

The Complete Agent & MCP Course — Udemy

Instructors: Ed Donner · Ligency
Completed: August 2026
Duration: 21 hours

Focused on:

• AI agents
• Tool use
• Agentic workflows
• Model Context Protocol
• Modern AI engineering patterns

Lahore University of Management Sciences — LUMS

Completed: May 2025

Focused on applied:

• Data Science
• Machine Learning
• Python
• Data analysis
• Model development

2026 → Present

Working across applied AI research, ML engineering and production-oriented AI systems.

Current areas:

Speech AI · LLM Systems · Computer Vision · Agentic AI · MLOps · AI Evaluation

2025 → 2026

Worked on analytics automation, experimentation and monetization analysis.

• Automated reporting workflows using Python
• Reduced a recurring analysis workflow from hours to seconds
• Built Power BI data pipelines
• Performed EDA across hundreds of datasets
• Designed and analyzed A/B tests
• Worked across multiple ad networks
• Contributed to measurable monetization improvements

Python
PyTorch
TensorFlow
Scikit-learn
Hugging Face
Whisper
YOLO
OpenCV

LangGraph
LlamaIndex
RAG
MCP
Ollama
Structured Outputs
Prompt Engineering

MLflow
Docker
Linux
CUDA
GitHub Actions
GitLab

FastAPI
Flask
PostgreSQL
SQL
REST APIs

Pandas
NumPy
Power BI
Matplotlib
Excel

I’m especially interested in systems where AI meets engineering.

Not just models that work in a notebook.

But systems that are:

measurable
deployable
observable
reproducible
testable
safe

Enough to operate in the real world.