AI / ML Engineer — Software Engineering Student

Salman
Khan.

I build intelligent software that learns, automates, and scales—from AI agents and LLM applications to production-ready ML systems.

0+Projects Shipped

0+ yrsMachine Learning

0%+Avg. Model Accuracy

View my work Get in touch

Scroll⌄

About

01 / Profile

I'm a final-year Software Engineering student at CECOS University, Peshawar, with two years of hands-on machine learning experience — mostly self-taught, all shipped. I care about the part after the notebook: turning a working model into a real, usable product.

My work spans computer vision, deep learning and — more recently — LLM-based automation and RAG pipelines. Off-screen: gaming, exploring, and driving.

Python Programming Certificate — upGrad
Solving Problems with Creative & Critical Thinking — IBM
Deep Learning: Getting Started — LinkedIn Learning

Languages

Python, C / C++, PHP

ML / DL

Supervised learning, classification, regression, CNNs, transfer learning, RAG & LLMs

Frameworks

TensorFlow, Scikit-learn, Pandas, NumPy, MediaPipe

Data & infra

SQL, MongoDB, FastAPI, Git

Automation

n8n workflows, LLM tooling, task automation

Human languages

English, Urdu, Pashto

Experience

02 / Timeline

AI Engineer — DevRolinAUG 2025 – JAN 2026

Worked as an AI engineer building with LLMs and RAG, and automating workflows with tools like n8n. Delivered 10+ projects, leading 3 of them end to end.

ML Intern — DeveloperHubMAY 2025 – AUG 2025

Completed ML/DL projects including house price prediction and heart disease detection, applying data analysis, feature engineering and fine-tuning to lift performance by 25%.

Python & Web Intern — P92 IT SolutionsOCT 2021 – AUG 2022

Contributed to 5 Python and 4 web development projects — feature implementation, debugging and performance work — while supporting decisions through data collection and market research. Gained hands-on practice with agile workflows, Git and code review.

Featured Project

03 / Final Year Project

Final Year Project · In Progress

AI Gym Trainer

A mobile fitness app that gives users personalized workout and diet plans, tracks progress, and corrects exercise form in real time — with an AI chatbot on hand for guidance along the way.

ProblemMost home workouts have no one watching your form — small mistakes in posture go uncorrected and turn into injuries over time.

How it worksCamera feed → pose landmarks → joint-angle calculation → rule-based form scoring → instant on-screen feedback, live while you train.

ArchitectureClient-server today: mobile frontend + FastAPI backend + MongoDB. The posture engine is being rebuilt as a standalone React Native module running a TFLite pose model fully on-device, so form correction works offline with zero network round-trip.

Computer Vision MediaPipe Pose FastAPI MongoDB React Native On-device TFLite

Source

More Projects

04 / Selected Work

01 / Detection System

AI Content & Plagiarism Detection System

A multi-modal academic-integrity engine — three detectors working together instead of one blunt score.

AI-text detection4-model ensemble (RoBERTa classifier, multi-model perplexity, stylometric & semantic analyzers), calibrated with isotonic regression to cut false positives.

Plagiarism searchFAISS semantic search + TF-IDF lexical matching across a 2.5M-sentence, 10GB corpus, with conditional live web search for suspicious lines.

Paraphrase detection5-signal weighted ensemble (semantic, structural, lexical, edit-distance, n-gram) to tell honest paraphrasing from disguised copying.

2.5M+Sentence Corpus

<10sPer Document

<5%False-Positive Target

RoBERTaFAISSFastAPI

02 / Deep Learning

Housing Price Prediction

Predicting house prices from more than just spreadsheet numbers.

ChallengeListing photos carry real pricing signal that tabular features alone miss — but images and numbers don't fuse easily.

ApproachFused a ResNet18 CNN over listing images with engineered tabular features in one multimodal model, tuned and evaluated on MAE/RMSE.

Result92% accuracy — a 13% lift over the tabular-only baseline.

TensorFlowResNet18CNN

03 / ML Pipeline

Titanic Survival Prediction

The classic ML problem, built the way production pipelines actually should be.

ChallengeMessy, partly-missing passenger data that needed to generalize cleanly to unseen records, not just fit a notebook.

ApproachScikit-learn Pipeline + ColumnTransformer for imputation, scaling and one-hot encoding, SelectKBest for feature selection, and GridSearchCV to tune a Random Forest.

Result+10% performance from feature selection alone; final pipeline exported and reusable via pickle.

Scikit-learnRandom Forest

04 / Computer Vision

Face Mask Detection

Real-time mask / no-mask detection through a live camera feed.

ChallengeRunning accurate classification fast enough for live video, not just static images.

ApproachTransfer learning on VGG16 with a custom binary classification head, integrated with MediaPipe for real-time face localization before classification.

ResultHigh-accuracy detection running live, ready for deployment on everyday hardware.

VGG16MediaPipe

05 / Healthcare ML

Heart Disease Detection

Predicting heart disease risk from clinical patient data.

ChallengeFinding real predictive signal in noisy, correlated clinical features without overfitting a small dataset.

ApproachCareful feature selection and preprocessing feeding a tuned Random Forest classifier.

Result88% prediction accuracy on held-out patient data.

Random ForestFeature Selection

1 / 5

Available for internships & full-time roles

Let's build
something.

Have a project, an internship opening, or just want to talk AI/ML? My inbox is open.

Email
salman@salmank.me Phone
+92 339 0030493

Location
Peshawar, Pakistan

SalmanKhan.