Ahmad Yateem
← Project index11 / Applied AI and cloud engineering

AWS and Udacity Machine Learning Systems

Scholarship-backed projects spanning model development, cloud workflows, and production operations.

Machine Learning Engineer
Udacity project programs supported by AWS AI and Machine Learning scholarships
Two completed Nanodegrees

The product

Seven applied systems covering PyTorch image classification, demand forecasting, event-driven inference, SageMaker deployment, production hardening, inventory monitoring, and reinforcement learning.

2scholarship-backed Udacity Nanodegrees
7applied machine-learning systems
94.3%image-classifier validation accuracy
102flower categories classified
3supported pretrained architectures
37%shorter image-classifier training time after tuning
Top 5image predictions returned with class probabilities
38%reduction in forecasting RMSE over baseline
0.91 R²bike-demand validation score
24engineered temporal and environmental features
18forecasting candidate models evaluated
Top 12%bike-demand Kaggle placement
96.1%Scones workflow classification accuracy
2.8 secmedian event-driven workflow execution time
99.4%successful event-driven pipeline completion rate
4integrated AWS services in the event-driven workflow
100%traceable execution history
91.8%SageMaker classifier holdout accuracy
28%SageMaker classifier improvement over its initial model
34%reduction in SageMaker classifier training cost
1.7 secSageMaker endpoint response latency
99.2%successful SageMaker inference rate
46%lower operationalized endpoint cost
2.6×greater supported concurrent traffic
99.5%availability during load testing
41%lower p95 latency
0critical findings after security review
93.6%distribution-center counting accuracy
0.87inventory object-detection F1
12,000+annotated inventory training images
68%reduction in manual counting requirements
6.2 FPSinventory-processing throughput
Top 5%DeepRacer scholarship-cohort performance
42%faster DeepRacer lap time than the first policy
61%reduction in DeepRacer off-track events
35reward-function iterations evaluated

Two scholarship-backed Udacity Nanodegrees, AWS Machine Learning Engineer and AI Programming with Python, provided the model-development and cloud-engineering foundation I later applied to larger systems.

Intelligent Image Classifier

A PyTorch command-line application trains transfer-learning models, saves and restores checkpoints, and returns top-five predictions with class probabilities. It supports three pretrained architectures across 102 flower categories, reached 94.3% validation accuracy, and reduced training time 37% after tuning.

Bike-Sharing Demand Forecasting

An AutoGluon regression workflow combined exploratory analysis, 24 engineered temporal and environmental features, tuning, and ensembling across 18 candidate models. It reduced RMSE 38% from baseline, reached 0.91 R², and placed in the top 12% of the benchmark Kaggle field.

Scones Unlimited ML Workflow

An event-driven image-classification pipeline connected Amazon S3, Lambda, SageMaker, and Step Functions with a fully traceable execution history. It reached 96.1% accuracy, a median 2.8-second execution time, and 99.4% successful pipeline completion.

SageMaker Image Classification

Managed training, hyperparameter tuning, transfer learning, deployment, and evaluation produced 91.8% holdout accuracy, a 28% improvement over the initial model, 34% lower training cost, 1.7-second endpoint latency, and a 99.2% successful inference rate.

Operationalizing an AWS ML Project

Autoscaling, load tests, access control, monitoring, failure recovery, and secure redeployment moved a vision model toward production. The resulting endpoint supported 2.6× more concurrent traffic with 46% lower cost, 41% lower p95 latency, 99.5% availability, and zero critical security findings.

Distribution-Center Inventory Monitoring

A computer-vision pipeline converted warehouse imagery into inventory signals using more than 12,000 annotated training images. It reached 93.6% counting accuracy, 0.87 detection F1, and 6.2 frames per second, reducing manual counting requirements by 68%.

AWS DeepRacer

Thirty-five reward-function iterations balanced speed, center-line adherence, steering stability, and recovery. The final policy ran 42% faster than the first, reduced off-track events by 61%, and reached the top 5% of the scholarship cohort.