Neural Network From Scratch

Context

Study repositories only help a portfolio when they show craft, not just curiosity. I kept this one because it moved beyond a notebook exercise into a small, testable system with deterministic evaluation and report artifacts.

Problem

I wanted to understand the mechanics of a neural network without hiding behind framework abstractions. At the same time, I wanted the repo to prove that learning projects can still honor engineering standards.

Constraints

No high-level ML framework.
Reproducible results for CI and portfolio demos.
Simple artifacts that can be inspected later.

Core proofDeterministic evaluation script

Repository signalLogs and smoke checks in CI

Operational concernPortable JSON and JSONL artifacts

Architecture

dataset utils
  -> normalization and split
  -> model
  -> backprop training
  -> evaluation metrics
  -> JSON / JSONL logs

Decisions and trade-offs

I used framework-free implementation because the learning value was the point.
I added a deterministic evaluation script because reproducibility is what separates a study repo from a throwaway experiment.
I kept logs in JSON and JSONL rather than adding a heavier experiment tracker too early.

What worked

The repo now communicates rigor instead of just enthusiasm.
The evaluation script provides an easy conversation path in interviews.
The CI-oriented smoke evaluation is stronger than a basic pytest -q alone.

What is still incomplete

A small benchmark report or confusion matrix export would strengthen the artifact story further.
There is room for richer experiment configuration without losing the repo's educational focus.

Evidence

Evaluation flowNeural Network From Scratch repository

Outputs:
- logs/eval-summary.json
- logs/eval-history.jsonl