I'm an AI/BI Engineer at Bluecascade and an active researcher in multi-modal learning, currently a 4th-semester BS Artificial Intelligence student (CGPA 3.81/4.0) at Emerson University, Multan. I build systems that don't just predict — they reason, explain, and generalize.

My work spans the full AI lifecycle: from novel architecture design and model training to MLOps pipelines, BI dashboards, workflow automation, and production deployment. I've contributed to PhD-level multi-modal AI research integrating vision, language, and audio — and I'm currently pursuing work on conflict-aware continual multi-modal learning for medical AI.

"The best model is not the one with the most parameters — it's the one that generalizes to the real world."

Full-time | Real-world production systems

• Architecting and deploying end-to-end AI and Business Intelligence solutions for clients
• Automated an intra-departmental decision-making pipeline using n8n + Airtable + JavaScript, slashing decision turnaround time from ~60 seconds to under 20 seconds — a >65% latency reduction in a live production environment
• Building BI dashboards and data flows that surface actionable intelligence from raw operational data
• Managing ML model deployment with MLOps practices including CI/CD integration and containerization

Independent | Advanced AI Research

• Contributed to a PhD-level research project focused on multi-modal learning combining vision and NLP
• Designed and implemented a novel Transformer-based architecture for dual-image medical report generation featuring:
  • Bilinear Cross-View Fusion — outer-product feature interaction between two image views for automatic importance weighting, replacing naive concatenation
  • Gated Cross-Attention Decoder — a learnable scalar gate (σ(w)) per decoder layer that dynamically balances self-attention vs. cross-attention contributions, allowing the model to learn how much to attend to image features vs. prior generated tokens
  • Three independent Transformer decoders (indication / findings / impression) sharing a single fused image encoder — enabling section-aware, clinically structured report generation
  • Dual decoding strategies — greedy decoding for speed and beam search (configurable beam size) for quality, evaluated against BLEU, METEOR, ROUGE-1/L, CIDEr, and RadGraph-F1 (clinical correctness metric)
• Integrated CheXNet (DenseNet121) pretrained on chest X-rays as the vision backbone with support for ImageNet weights, custom CheXNet weights, and fine-tuning modes
• Implemented Grad-CAM explainability for visual grounding of model attention on X-ray images
• Used Bio_ClinicalBERT tokenizer for clinical text tokenization, with masked loss and masked accuracy to handle variable-length padded sequences correctly

• Built ML/DL pipelines on real-world datasets covering financial anomaly detection, medical image cancer classification, and NLP-based text classification
• Deployed models via Flask REST APIs and optimized data preprocessing pipelines for production throughput
• Worked on stock market anomaly detection using time-series modeling

• Built and deployed ML models for real-time restaurant data analysis
• Improved accuracy of recommendation systems using regression and classification algorithms with feature engineering on structured tabular data

• Developed a QA chatbot and language translation models using NLP pipelines
• Created a prompt-to-music generation tool leveraging generative AI
• Implemented real-time object detection using YOLO architecture

Investigating how multi-modal AI models (vision + language) can learn sequentially from new medical data without catastrophically forgetting prior knowledge — with a specific focus on conflict detection between modalities during continual learning. This targets a core open problem in clinical AI: how to keep models up-to-date in dynamic hospital environments without full retraining.

Multi-Modal Deep Learning | Medical AI | NLP

A full Transformer-based system that takes two X-ray views (frontal + lateral) and generates structured radiology reports (indication, findings, impression) with clinical correctness evaluation.

Architectural novelties:

• Bilinear cross-view image fusion (outer-product interaction)
• Gated cross-attention with per-layer learnable α gate
• Triple-decoder architecture for section-aware generation
• Evaluated with RadGraph-F1 for entity-relation clinical accuracy

TensorFlow Bio_ClinicalBERT CheXNet Grad-CAM Beam Search RadGraph

VSCode Extension | NLP | Developer Tools

A VSCode extension that reviews code in real-time using Hugging Face models for linting suggestions and optimization hints — bringing AI pair-programming directly into the editor.

JavaScript Python VS Code API Hugging Face NLP

Flask Web App | NLP | Education AI

An interactive web application that solves and explains math problems from basic arithmetic to advanced calculus, and a chatbot interface.

Flask HTML Data Visualization NLP Transfer Learning

• 🔬 Conflict-Aware Continual Multi-Modal Learning — medical AI research (active)
• 🚀 Releasing new open-source projects on GitHub soon
• 📦 Deepening MLOps practices — model versioning, monitoring, and deployment pipelines
• 🧩 Exploring agentic AI systems using LLMs with tool-use and memory

"Building AI that understands the world — one modality at a time."

⭐ If you find my work interesting, consider starring a repo — it means a lot!