DEV Community: Furkan

I built a from-scratch Transformer + MiniGPT in pure Python (no PyTorch/TF/NumPy) to learn how it all fits feedback on the autograd?

Furkan — Sun, 28 Jun 2026 12:10:07 +0000

Cognitive Discovery System (CDS) a scientific computing library written in pure Python. No NumPy, no SciPy, no compiled extensions. Zero dependencies, runs anywhere.

18 modules: quantum circuit simulation, FFT/IFFT, ODE solvers, numerical integration, linear algebra, statistics, Monte Carlo, optimization, graph algorithms, symbolic math, a from-scratch NLP stack (BPE → attention → MiniGPT + scalar autograd), and a hypothesis-generation module.

The idea: pure Python can't beat C on raw loops, but it can close the gap algorithmically — O(N log N) FFT, O(N³) LU decomposition (instead of O(N!) determinants), O(1) probabilistic quantum measurement. Not a NumPy replacement; a readable, dependency-light toolkit.

1,441 tests · 100% coverage · on PyPI https://pypi.org/project/cognitive-discovery-system/

Repo: https://github.com/Furox88/cognitive-discovery-system Install: pip install cognitive-discovery-system

If you find it useful, a star on the repo means a lot. The autograd part is the one I'd most appreciate eyes on.
If anyone has time to glance at the from-scratch autograd implementation, that's the part I'm least sure about. Open to feedback

I Built a Pure Python AI Library with Zero External Dependencies – No NumPy, No TensorFlow

Furkan — Sun, 21 Jun 2026 11:37:08 +0000

Hi everyone,

I wanted to share a project I've been working on for the last 12 days. It's called Cognitive Discovery System (CDS) — a scientific computing library written in pure Python with zero heavy dependencies. No NumPy, no SciPy, no compiled extensions.

Why does this exist?

I'm not trying to compete with NumPy — it can't, and it shouldn't try. The goal was:

Readable source you can actually learn from (every algorithm from scratch)
Edge runtime — runs anywhere Python runs, no BLAS/C-Fortran toolchain
One umbrella for math + physics + stats + ML + signals + quantum, instead of 6+ separate libraries

What's inside (17 modules)

Quantum — single & multi-qubit circuits, Bell/GHZ states, entanglement detection
Signals — O(N log N) FFT, convolution via Convolution Theorem, power spectrum
Optimization — gradient descent, Newton's method, Adam
Numerical — O(N³) LU/QR/Cholesky, RK4 ODE solver, Gauss-Legendre quadrature
Stats — t-test, chi-square, ANOVA, Pearson, linear regression
ML — neural network from scratch (MLP, Adam)
NLP — BPE tokenizer, multi-head attention, MiniGPT demo
Graphs — Dijkstra, Kruskal MST, topo sort
+ Monte Carlo, symbolic math, knowledge graph, hypothesis generation

One design choice I'm proud of

The whole library is built around "smarter algorithms, not faster loops." For example:

Instead of naive shot-by-shot quantum sampling, I used O(1) probabilistic sampling with explicit state collapse
The linear algebra uses partial-pivoting LU (O(N³)) instead of naive determinant expansion (O(N!))
FFT uses zero-padded radix-2 (O(N log N)) instead of naive DFT (O(N²))

Engineering

1,244 tests, 100% statement + branch coverage
Multi-OS CI matrix, CodeQL, signed PyPI releases (OIDC)
mkdocs site, CLI, interactive web dashboard

Install

pip install cognitive-discovery-system

Repo: https://github.com/Furox88/cognitive-discovery-system
Docs: https://Furox88.github.io/cognitive-discovery-system/

Feedback wanted

Is the "pure Python, no NumPy" pitch compelling, or does it sound like reinventing the wheel?
The library tries to be a broad umbrella — is that a strength (one package) or a weakness (unfocused)?
Any modules where the from-scratch implementation would be useful for learning?

Not trying to compete with NumPy — genuinely curious whether there's a niche for a readable, dependency-light alternative. Happy to answer questions about any of the implementations.

Building a Scientific Computing Platform (Quantum, ML, Math) in Pure Python Without NumPy

Furkan — Mon, 15 Jun 2026 16:54:45 +0000

Hi everyone,

We rely so heavily on tools like NumPy, SciPy, and PyTorch that most of us treat them like black magic. I wanted to break that illusion. My goal was to
see if I could build a professional-grade computational science platform using nothing but the Python Standard Library.

It turned into an obsessive deep dive into algorithmic optimization, memory management, and raw math.

I've just released v0.6.0, and here is what it currently handles natively in Pure Python:

Linear Algebra & Calculus: I had to scrap naive 0(N!) determinants for an O(N3) Partial Pivoting LU decomposition to stop my machine from melting. Also
includes an Adaptive Newton Method.
Quantum Mechanics: A complete state-vector simulator. Instead of calculating matrices thousands of times for measurements, I implemented an 0(1)
probabilistic sampling algorithm with true state collapse.
Machine Learning from Scratch: A fully functioning Neural Network (MLP) trained by a custom-built Adam optimizer that features full momentum state
persistence.
Signal Processing: An O(N log N) zero-padded 1D/2D FFT and Convolution Theorem implementation.
Statistics: An autonomous hypothesis engine that takes generated ideas and tests them against datasets using ANOVA and Welch's t-tests.

The Ecosystem Upgrade:
To prove it wasn't just a toy, I made sure the engineering standards were flawless:

350+ unit tests hitting 95% coverage.
Strict typing enforced by mypy and ruff in the CI pipeline. Interactive Jupyter Notebook tutorials and a fully automated MkDocs website.
A Streamlit dashboard and a Rich-powered CLI.

If you are a student who wants to read the actual math behind an FFT or a Quantum Gate without getting lost in C/C++ source code, or if you're just
curious about how far vanilla Python can be pushed, I'd love for you to explore it. Feedback, code reviews, and stars are highly appreciated!

GitHub Repo: https://github.com/Furox88/cognitive-discovery-system

Documentation & Tutorials: https://furox88.github.io/cognitive-discovery-system/

PyPI Installation: pip install cognitive-discovery-platform