evals

Practice exercises for VitalView AI interns. This is where new engineers get evaluated on their first tasks: small, scoped exercises that mirror how we actually work, not toy problems disconnected from real practice.

What this is

Each exercise lives under exercises/ with its own README describing the objective, constraints, and acceptance criteria. The code itself is usually small. The point is the process: branching, commit hygiene, tests, and a PR that tells a clear story.

How to submit a solution

We use a fork + pull request workflow:

1. Fork this repository.
2. Clone your fork and follow the branching and commit standards in CONTRIBUTING.md.
3. Work the exercise in solutions/<your-name>/<exercise-name>/ in your fork.
4. Push your branch and open a pull request back into VitalView-AI/evals main.
5. Request a review.

Solutions are reviewed like real PRs. Once merged, they stay in the repo so other interns can see different approaches.

Exercises

1. exercises/nn-xor — implement a feedforward neural network from scratch in NumPy that learns XOR.
2. exercises/mnist-classifier — extend it to a multi-class handwritten digit classifier (softmax, mini-batch SGD, train/test split).
3. exercises/cnn-from-scratch — extend it again to a convolutional network (hand-implemented conv2d/maxpool), compared against the level-2 model on parameters, accuracy, and training time.

Ground rules

• Keep your branch scoped to the exercise only.
• AI coding assistants (Claude Code, etc.) are welcome, but you own everything you submit. If a reviewer asks you to explain a line, you should be able to.
• See CONTRIBUTING.md for full git, commit, and environment standards.