Laiba Azhar — AI/ML Engineer

About

I'm a 7th-semester Artificial Intelligence student at COMSATS University Islamabad (CGPA 3.60/4.0), building toward AI/ML engineering roles. Most of my work sits at the uncomfortable edge where machine learning meets production reality: models that hallucinate, pipelines that break silently, retrieval systems that confidently return the wrong answer.

My defining project is a medical insurance claims processing system where a wrong answer isn't a bad user experience, it's a wrong claim. That constraint shaped how I think about AI engineering: build in verification, not just capability.

Outside coursework, I build full end-to-end systems: FastAPI and React on the app layer, self-healing MLOps underneath, and retrieval-augmented generation for domain-specific knowledge.

100%

precision on RAG eval set

5,533

vector chunks indexed

layers of hallucination defense

3+ yrs

teaching CS/ML internationally

Experience

AI/ML InternJul 2026 · 6 weeks

AXI Technologies

Six-week AI/ML internship focused on applied machine learning work in a production engineering environment.

Artificial Intelligence InternJul – Aug 2025

e-strats · on-site, Islamabad

Built machine learning pipelines for data preprocessing, feature engineering, and model evaluation. Worked end-to-end on classification and regression tasks with real production datasets.

Marketing LeadApr – Aug 2025

CAIS · COMSATS AI Society

Led marketing and outreach for a 200+ member university AI society, coordinating events and campaigns that grew attendance and engagement.

FounderFeb 2025 – present

Alif Relief Organization

Built organizational strategy and volunteer management systems for community support initiatives, coordinating multi-team operations and fundraising.

Part-time TutorFeb 2025 – present

IbaTutoring · Remote

Designed machine learning curricula covering Python, supervised learning, model evaluation, and algorithm design for international students.

Computer Science InstructorAug 2020 – present

Preply · Part-time, remote

Taught Python, Java, and C++ to university students, including at the University of Greater Manchester and University of Melbourne, across 16+ skill areas with adaptive teaching methods.

Community & activities

Founder — Alif Relief Organization

Google Developer Groups — Data science lead

CAIS-COMSATS AI Society — Marketing lead

ILC COMSATS — HR

Projects

Flagship project

Multistage Wound Screening AI

96.3% detection accuracy

binary wound detection

→

9-class classification

→

OOD robustness check

Why multi-stage

Single-model wound classifiers fail silently on inputs they weren't trained for, a burn misclassified with false confidence, or a non-wound image processed as if it were one. This pipeline splits the problem into three stages so each failure mode gets caught explicitly instead of hiding behind one softmax output.

The pipeline

Stage 1 filters non-wound images: 5 architectures benchmarked (CustomCNN, MobileNetV2, ResNet50V2, EfficientNetB0/B3) on 6,705 images, MobileNetV2 won at 96.3% accuracy and 0.994 AUC. Stage 2 classifies detected wounds into 9 clinical categories (bruise, pressure ulcer, infected wound, abrasion, laceration, venous ulcer, burn, cut, diabetic ulcer) across 746 images with 5-fold cross-validation, EfficientNetB3 led at 71.3% CV accuracy. Stage 3 stress-tests all 5 models against 233 out-of-distribution images to measure how often the system over-trusts inputs it should flag as uncertain.

Edge deployment

Scored every model on a composite of accuracy, inference speed, and size to pick a deployment target. MobileNetV2 scored highest (0.908) and was exported to TFLite for mobile and edge use, trained via DirectML so it runs without a CUDA GPU.

Known limitation, documented honestly

Stage 3 OOD detection didn't hit the target false-positive rate (under 10%) on any architecture, the best came in at 66.5%. That means the system currently over-trusts out-of-distribution inputs more than a real clinical deployment would allow. Logged here as the clearest next research direction rather than hidden: confidence calibration, a dedicated OOD head, or Mahalanobis-distance-based rejection.

TensorFlow / KerasMobileNetV2EfficientNetGrad-CAMTFLiteDirectML

Source Built with Urva Ishfaq

Medical Billing RAG System

100% precision

264 articles chunked

→

5,533 vectors indexed

→

grounded, cited answer

A production RAG system that processed 264 medical articles into 5,533 vector chunks, reaching 100% precision on the evaluation set with full source attribution and sub-second retrieval. Semantic chunking (800 characters, 150-character overlap) preserved medical context, sentence-transformers/all-MiniLM-L6-v2 handled embeddings, and Mistral 7B via the Hugging Face Inference API served every answer with a required source citation. Chose the smaller Chroma plus MiniLM stack over Pinecone/Weaviate and larger embedding models after benchmarking showed no precision gain, just added latency and cost, for local deployment.

ChromaDBMistral 7BLangChainsentence-transformersGradioHugging Face Spaces

Live demo Source

AI-Powered Medical Insurance Claims Processing

deployed

ocr extraction

→

3-layer validation

→

claim decision

A full-stack claims processing system where Groq handles OCR extraction from claim documents and a three-layer hallucination-detection safety net catches and flags errors before they reach a decision. Built to treat AI confidence as a signal to verify, not a signal to trust.

FastAPIReactGroqRailwayVercel

Live demo Source

Self-Healing MLOps Pipeline

in progress

monitor

→

detect drift/failure

→

auto-recover

An MLOps pipeline designed to detect model or infrastructure failures on its own and recover without manual intervention, using CI/CD automation and containerized rollback strategies.

JenkinsDockerKubernetesAWS EC2

Source