Madiha Iqbal

AI Engineer & Researcher

Building and evaluating AI systems through computer vision, generative modeling, and evidence-grounded LLM applications.

I worked across the machine learning lifecycle, from model development and experimentation to evaluation and practical AI applications.

• Current Focus: Deep learning for object detection and semantic segmentation; research manuscript in progress.
• Research Interests: Synthetic data generation, multimodal AI, LLM alignment, evidence-grounded RAG, and trustworthy AI.
• Goal: Building reliable, measurable, and useful AI systems through rigorous experimentation.

Deep Learning for Object Detection and Semantic Segmentation

Research on deep learning pipelines for detecting and segmenting structural defects in concrete.

• Trained and evaluated DeepLabV3 and SegFormer models on structural-defect datasets.
• Developed a four-class U-Net segmentation pipeline with custom dataset loading and checkpoint management.
• Applied YOLOv8 for crack, spalling, and exposed-rebar detection.
• Evaluated model performance using class-wise IoU, F1-score, precision, and recall.

Demo: The manuscript is currently in preparation. Therefore, the source code and full implementation are not publicly available on GitHub. The repository will be made public after publication.

Tech Stack: PyTorch,SegFormer, DeepLabV3, U-Net, DDRNet, YOLOv8, yolov26,OpenCV.

Exploring synthetic knee X-ray generation to address medical-data scarcity and privacy challenges.

• Trained StyleGAN3 on 512 × 512 knee radiographs.
• Fine-tuned Stable Diffusion 1.5 using LoRA for domain-specific image generation.
• Experimented with pix2pix for image-to-image translation.
• Evaluated generated images using FID, SSIM, PSNR, and LPIPS.

Tech Stack: PyTorch, StyleGAN3, Stable Diffusion, LoRA, pix2pix.

An evidence-grounded medical question-answering system designed to provide answers based on retrieved medical documents.

• Built a Retrieval-Augmented Generation pipeline using document chunking, semantic embeddings, and FAISS vector search.
• Integrated local LLM inference with Ollama and Llama 3.2.
• Evaluated answer faithfulness, relevancy, and context precision using RAGAS.
• Reported evaluation results of 96.7% faithfulness, 91.2% answer relevancy, and 80.1% context precision on a curated medical QA benchmark.

Tech Stack: Python, LangChain, FAISS, Ollama, Llama 3.2, RAGAS.

Experimenting with preference alignment to improve small language models for medical question answering.

• Fine-tuned Llama-3.2-3B-Instruct using LoRA-based Supervised Fine-Tuning and Direct Preference Optimization.
• Designed pairwise, LLM-judged evaluations across accuracy, completeness, safety, clarity, and helpfulness.
• Observed a 90% preference rate for DPO-tuned outputs in head-to-head comparisons against SFT outputs in the evaluated experiment.

Tech Stack: Hugging Face Transformers, TRL, PEFT, LoRA, DPO.

An AI-based inventory monitoring and decision-support system for stock-risk classification and reorder recommendations.

• Developed inventory monitoring logic for low-, medium-, and high-risk stock levels.
• Built a FastAPI backend and Streamlit dashboard for inventory insights.
• Explored agent-based workflows and automated inventory decision support.

Tech Stack: Python, FastAPI, Streamlit, Pandas, Scikit-learn.

A local AI assistant combining conversational LLMs, document retrieval, and voice interaction.

• Integrated local LLM inference using Ollama.
• Built a document-based RAG workflow for contextual question answering.
• Developed speech-to-text and text-to-speech components for voice interaction.

Tech Stack: Python, Streamlit, LangChain, Ollama, FAISS, FFmpeg.

• LinkedIn: Madiha Iqbal
• Email: madihaiqbal786786@gmail.com