4 *Real* Machine Learning Projects That Get You Hired - No More Tutorials!

Summary

The video argues that AI/ML portfolios are stronger when they showcase self-directed, end-to-end projects rather than tutorial reproductions.

Project planning guide

1. Choose a problem worth solving, even if it is small or playful.
2. Sketch a simple system design: identify data sources, storage, model approaches, and how users will access the results.
3. Build a working MVP quickly, then improve it rather than trying to design a perfect system upfront.

Project examples

1. Shelf Scanner — book discovery app

Users share their reading preferences and Goodreads data, then photograph a bookshelf. The app identifies the books and makes personalized recommendations.

• Technology: PostgreSQL, Multer for image uploads, GPT-4o Vision for book-spine recognition, and Google Vision as a fallback. GPT-4o generates recommendations.
• Product: A mobile-friendly React/Express app deployed on Vercel.
• Additional features: Metadata caching, rate limiting, monitoring, an admin dashboard, tests, and security measures.
• What it demonstrates: The practical AI product lifecycle, including API fallbacks, cost control, and deployment.

2. PackVote — group travel planning app

The app collects trip preferences through Twilio SMS surveys, recommends destinations, and lets group members rank options to reach a consensus.

• Technology: A model gateway routes tasks to OpenAI, Anthropic, or DeepSeek.
• Additional features: Versioned prompts, prompt A/B testing, evaluation metrics, usage and cost tracking, Prometheus monitoring, rate limiting, and integration tests.
• Status: Described as an intermediate-level project that was not yet fully deployed. A planned improvement is an agent that retrieves real-time pricing.

3. “Are You a Cat?” — MLOps classification system

This humorous image-classification project focuses on production ML operations rather than developing a novel model.

• Model: A standard TensorFlow CNN trained on cat, dog, selfie, and miscellaneous image datasets.
• Workflow: ZenML orchestrates training, deployment, and inference pipelines; MLflow tracks experiments and model artifacts.
• Validation and deployment: Deepchecks validates data. Deployment proceeds only if the model meets precision and recall thresholds.
• Feedback loop: A Streamlit interface serves predictions and saves user feedback and images to S3 for potential future training and drift detection.
• What it demonstrates: Reproducible training, validation, deployment automation, monitoring, and feedback-loop design.

4. Days to Banana Death — computer-vision regression project

This proposed project predicts how many days remain before a banana becomes overripe.

• Dataset: Photograph bananas daily and label each image with the number of days remaining.
• Model: Fine-tune a pretrained model, such as ResNet, to predict a single numeric value.
• Suggested implementation: A camera-enabled web app, a lightweight FastAPI backend, and a mobile-friendly interface that could display a confidence interval.
• What it demonstrates: Data collection, image modeling, regression, and end-to-end application development.
• Status: Presented as a project concept, not a completed build.

Additional mention

The video briefly promotes Akiflow as a tool for planning a week, consolidating tasks, time-blocking work, and tracking time across projects.

Main speaker and sources

• Main speaker: Marina Wyss, sharing project examples and advice based on her own work and mentee projects.
• Other credited source: A mentee at UC Berkeley is credited with the banana-project idea.