Abstract

Real-time Dashboard: Offers insights through an interactive, real-time dashboard.

Analysis of Historical and Current Data: Examines and tracks both past and present wildfire incidents.

Utilization of Multiple Data Sources: Leverages real data from Alberta’s Historical Wildfire Data and Natural Resources Canada for comprehensive analysis.

Focus on Data Accuracy and Currency: Ensures accuracy and currency of data throughout the project’s development which is critical for making informed decisions.

Real Data: Using real open city data.

Scalability: Pipeline deployed in Docker which contributes to scalability and reproducibility.

Preprocessing: Preprocessing the data so it conforms to the data warehouse schema. This includes cleaning, dropping columns, etc. (Data cleansing/sanitation)

Data Management Efficiency: Deployed PgAdmin instance to manage ingested data.

Snap & Organize: Capture receipts and bills in-app, ditching paper clutter for organized records.

Visualize Spending: Understand your finances at a glance with WalletWatch’s visual spending insights.

Timely Alerts: Receive notifications to stay informed and in control of your financial status.

Prototyping & User Feedback: Refined the app through paper prototypes, focus group testing, and wireframe mockups.

Collaboration: Worked in a team of 5 to build and manage various aspects of the app, including user testing.

Contribution: I was responsible for overall UI design, the form to add Expenses, and photo management (uploading, database storage, deletion).

Deployment: A working model in a Jupyter notebook is only 10% of the battle. I focused on deploying this model to a Dockerized production environment for scalability, reliability, reproducibility.

Model Versioning: Model parameters were saved and versioned to provide model versioning

NB Classifier: building a Naive Bayes classifier from scratch to demonstrate robust understanding of the classifier.

Model Evaluation: With a balanced dataset of exactly 50% positive and 50% negative class (feeds into the Naive prior probability) and a custom NB model, we get weak results. Of course major improvements can be made by leveraging more powerful models that don’t have the inherent naivety of prior probabilities.

SWE: Taking this project farther by implementing an interactive avatar with animated emotions matching the classified sentiment.

Preprocessing: Using NLTK to remove stopwords, sanitizing input, text normalization.. etc.

Customer Value: Customers that stay longer tend to pay more monthly.
Significance: Businesses should prioritize brand loyalty by rewarding long-term customers with incentives like points, special offers, as they tend to have higher monthly spending.

Customer Retention: New customers are the most likely to leave.
Significance: Businesses should focus retention efforts on new customers.

Model Evaluation: Random Forest algorithm is used in a “grey-box” approach, achieving more explainability and control in contrast to the “black-box” nature of some commercial tools while having slightly better performance.

Data Source: Real-world dataset from IBM, representing telecommunications customer data, which still needed to be preprocessed