How often do you commit?

Daily since 1 July 2026 — the mission is 548 days to 31 Dec 2027. Each day is a notebook.

Do you use PyTorch or pure NumPy?

Both — I first build everything in pure NumPy by hand (backprop, Adam, CNN), then show the PyTorch version for comparison.

Can I follow along locally?

Yes — clone hariomlohardev/AGI_Research, open any notebook in Jupyter, and run top to bottom. No API keys needed.

Hariom Lohar — AGI_Research · Daily log

A public lab notebook committed daily since 1 July 2026 — 548 days to 31 Dec 2027. I derive everything by hand: linear algebra, calculus and probability from axioms, then neural nets, backprop, optimisers from SGD to Adam in pure NumPy, now CNNs and Transformers.

Each day is a notebook with math, code and reflection — no autograd, no shortcuts. The repo is the research log; the portfolio shows the distilled logs. Follow along as the bench grows.

Highlights

Linear algebra, calculus and probability from axioms
Forward and backward pass in pure NumPy
Backpropagation derived by hand — no autograd
Optimisers from SGD to Adam, built step by step
Now: CNNs, attention, Transformers

backprop ✓ NumPy CNNs Transformers PyTorch daily commits

Jupyter Notebook 86.4% · Python 13.6%

Days

40+

Notebooks

40+

Commits

150+