NERV is an educational, applied machine learning project that demonstrates how trained ML models move from experimentation into structured, real-world applications.
It focuses on the often-skipped middle layer of ML learning:
bridging model training → validation → deployment → inference.
This repository is not about chasing accuracy benchmarks.
It is about using trained intelligence correctly.
NERV is:
- A learning-oriented ML system
- A reference for applying trained TensorFlow models
- A practical example of ML + web integration
- A documented pipeline from training artifacts to inference
NERV is NOT:
- A production ML framework
- An AutoML tool
- A plug-and-play library
If you are learning how ML systems actually live beyond notebooks, this project is for you.
NERV is intentionally split into two layers:
All model training, preprocessing, and evaluation live here:
🔗 Training Repository (TensorFlow-focused)
https://github.com/aypy01/tensorflow
This includes:
- Data preprocessing
- Model architectures
- Training strategies
- Evaluation metrics
- Saved
.kerascheckpoints
NERV consumes those trained models and:
- Loads them as versioned artifacts
- Integrates them into a Django application
- Runs controlled inference
- Demonstrates real-world usage patterns
This separation mirrors how ML systems are structured in practice.
| Model | Task | Dataset | Accuracy |
|---|---|---|---|
titanic.keras |
Binary classification | Titanic Survival | ~81% |
iris_species.keras |
Multiclass classification | Iris Dataset | ~70% |
cifar10.keras (Oculus) |
Image classification | CIFAR-10 | ~72% |
sentiments.keras (Yapper) |
Text classification | IMDb Reviews | 85.80% |
📁 Model files are stored in:
https://github.com/aypy01/nerv/tree/main/nerv/models
Models are intentionally consumed through code, not as standalone downloads.
-
Titanic Survival Prediction
Classical tabular ML workflow with feature engineering. -
Iris Species Classification
Multiclass classification using dense networks. -
Oculus (Computer Vision)
CNN-based image classification on CIFAR-10. -
Yapper (NLP)
Sentiment analysis using embeddings and BiLSTM.
Each component demonstrates a different ML modality while following the same deployment discipline.
Full Project Documentation:
https://aypy01.github.io/docs/nerv/nerv.html
The documentation explains:
- End-to-end training → inference flow
- Model design choices
- Integration decisions
- Common ML mistakes avoided
Think of the docs as a guided walkthrough, not just reference material.
- TensorFlow / Keras
- Python 3
- Django
- HTML / CSS
- JavaScript
- Kiran Jain - My primary school homeroom teacher, whose early trust and encouragement shaped my confidence in learning and building.
- David J. Malan CS50 instructor
- Brian Yu CS50 Web
- TensorFlow Model training & inference
- Django Backend & web integration
- CS50 Computer science foundations
- scikit-learn Classical ML utilities
- Google Colab Experimentation & prototyping
- ChatGPT Debugging, documentation
- The Motivation
NERV was born during a period of professional transition. After leaving my previous role, I decided to bridge my knowledge from CS50 AI and CS50 Web. I realized that while many people can train a model, very few know how to give that model a "home" in a real application. I built this to challenge myself: to take "invisible" Python scripts and turn them into a visible, interactive system. If this project helps even one learner understand how to connect AI to the web, then the mission is accomplished.
NERV is not about experimenting blindly.
It is about applying trained intelligence correctly.
Created and maintained by
