Umer Haroon

Electrical Engineer specializing in AI Automation, Computer Vision, and LLM Systems.

Based in: Islamabad, PKMost recently: AI/Data Science Intern, Systems Limited

Electrical engineering graduate with hands-on experience in building end-to-end AI systems, including RAG-based automation and real-time computer vision pipelines. Skilled in Python, FastAPI, and PyTorch, he excels at developing practical solutions from model optimization to containerized deployment.

Experience

AI/Data Science Intern

Systems LimitedIslamabad, PKinternshipJun 2025 – Aug 2025

Summary

Asynchronous APIs & Workflow Automation: Built high throughput model serving endpoints using FastAPI and engineered automated, event driven data ingestion pipelines using n8n webhooks and REST APIs. Inference Optimization & Deployment: Optimized deep learning and transformer pipelines using PyTorch and ONNX Runtime, significantly reducing inference latency for real time applications. Semantic Retrieval & Data Processing: Implemented vector based semantic search and automated feature extraction workflows to handle unstructured datasets and feed downstream services. Containerization & Automated Testing: Containerized internal microservices using Docker and built automated testing scripts to benchmark latency, throughput, and accuracy across dataset updates.

What I did

Built automated benchmarks to objectively compare latency, throughput, and accuracy across model versions.

Results

Significantly reduced inference latency for real time applications through optimization

FastAPIn8nREST APIsPyTorchONNX RuntimeDockerTransformersVector Search

Skills

Technical

Python

Computer Vision

Microsoft Office Suite

LangChain

RAG

Pandas

NumPy

n8n

FastAPI

PyTorch

LLMs

REST APIs

OpenCV

YOLO

Streamlit

Git

GitHub

Canva

Scikit-Learn

Postman

Gemini API

NLP

Gradio

Docker

SQL

C++

Django

Multinomial Naive Bayes

BLEU

Logistic Regression

OCR

Feature Engineering

BLIP

ONNX Runtime

Projects

AI Support Ticket Automation System

Lead Developerpersonal

Built an AI powered support ticket automation pipeline using FastAPI, Streamlit, Gemini API, LangChain RAG, and n8n. Implemented structured LLM based ticket analysis for category, priority, sentiment, escalation status, summary generation, and suggested replies using Pydantic schemas. Added LangChain based retrieval over local support policy documents using semantic search to ground LLM responses in company specific rules. Integrated n8n production webhooks with IF based routing, Google Sheets logging, and Gmail alerts for escalated tickets.

FastAPIStreamlitGemini APILangChainRAGn8nPydanticGoogle SheetsGmail API

Industrial Visual Anomaly Detection System

Lead Developerpersonal

Built an end to end industrial anomaly detection system using PyTorch and pretrained ResNet-18 patch features. Evaluated the system across all 15 MVTec AD categories for image level defect classification and pixel level anomaly localization. Generated anomaly heatmaps and binary defect masks to visually explain predicted defective regions. Implemented automatic category detection, CLI inference, and a Streamlit web application for interactive testing.

PyTorchResNet-18MVTec ADStreamlitCLI

Real-Time Vehicle Detection and Speed Estimation System

Lead Developerpersonal

Developed a vision based roadside speed monitoring pipeline using YOLOv4-tiny, SORT tracking, and Lucas-Kanade optical flow. Implemented homography based camera calibration to convert pixel displacement into real-world distance for speed estimation. Deployed the system on NVIDIA Jetson Nano, using frame downsampling and feature-point reduction for real-time operation. Generated annotated output videos containing vehicle IDs, bounding boxes, tracked feature points, and estimated speeds.

YOLOv4-tinySORT trackingLucas-Kanade optical flowNVIDIA Jetson NanoOpenCV

Image Captioning System Using Deep Learning

Lead Developerpersonal

Implemented a multimodal vision to text pipeline using PyTorch and the Salesforce BLIP (Vision Encoder + Text Decoder) architecture. Benchmarked greedy decoding against beam search generation on the Flickr8k dataset, quantitatively evaluating output quality across BLEU-1 through BLEU-4 metrics. Built and deployed an interactive Gradio web application for real time custom image uploads and caption generation.

PyTorchSalesforce BLIPFlickr8kGradioBLEU metrics

Urdu News Classification System

personal

Built an end-to-end Urdu news classification system using OCR, text preprocessing, and multiple ML models to address challenges of building NLP systems for a lower-resource language.

OCRNLPscikit-learnTF-IDFLogistic RegressionMultinomial Naive BayesFlaskLSTM

Education

Bachelor of Electrical Engineering

NUSTElectrical EngineeringGPA 3.22026

Coursework

Deep LearningMachine LearningComputer VisionEmbedded System DesignDigital System Design

Hire UmerReach out about a role, a contract or a conversation.For recruiters

Umer's twin is AI, it can make mistakes.

More people on Kanvis

Browse more people Create your own Kanvis