Self-taught from India. No bootcamp, no shortcut — math from axioms → neural nets & backprop in NumPy → CNNs, Transformers & PyTorch. All in public.
AGI Mission — 1 July 2026 → 31 Dec 2027 · 548 days · 8h/day · derive before import
Day 040 — multi-head attention + residual + layernorm by hand (NumPy). Yesterday: scaled dot-product softmax(QKᵀ/√dₖ)V + causal mask. Every day committed to AGI_Research.
- Now: CNNs, attention & PyTorch internals · Daily commits · Live benches
- Next: Full Transformer block → LLM → RAG at scale
- How:
F3F0E8paper,FFD400signal,0B1220ink — same Lab Notebook onhariomlohardev.github.io
▸ Mission ruler — 4 phases
| Phase | Title | Detail | State |
|---|---|---|---|
| 01 | Math foundations | linear algebra · calculus · probability | ✅ done |
| 02 | Neural nets from scratch | backprop · optimizers · NumPy | ✅ done |
| 03 | Deep learning | CNNs · Transformers · PyTorch | ● now |
| 04 | Toward AGI | LLMs · RAG at scale · research | ○ next |
Progress bar & countdown live on the site: #mission · 8h/day · open notebook.
New post = new file in
posts/*.md→node scripts/generate-blog.js→blog/p/<slug>/+feed.xml. Zero $, zero CMS.
- Day 040 — Multi-Head, Residual & Layernorm by Hand —
2026-08-10· open → ·daily-log / Transformers / NumPy - Day 039 — Attention by Hand —
2026-08-08· open → ·daily-log— derivedQKᵀ/√dₖ+ mask + softmax Jacobian - Spam Classifier Shipped — 5,572 Messages —
2026-08-05· live bench → ·Python / Naive Bayes / Pyodide — runs in your browser - More → hariomlohardev.github.io/blog.html · RSS → feed.xml · Newsletter → buttondown.com/hariomlohar (free)
| ✅ | Math Foundations & Autograd | Hand-derived SVD, eigendecomposition, PCA · Jacobians & chain rule · autograd engine · Neuron → Layer → MLP on top · custom Adam with a tweak of my own |
| ✅ | Neural Nets — NumPy | Backprop derived by hand · SGD → Adam · validated on `T=3,d=4` numeric grad 1e-4 — live in AGI_Research Day 001 → 039 |
| ● | Probability, Info Theory & Classical ML | Naive Bayes (shipped, 5,572 msgs) · Decision Trees · Random Forests · SVMs · K-Means — each from scratch, checked vs scikit-learn |
| ● | Deep Learning & Transformers — now | CNNs & RNNs by hand → PyTorch · Transformers · multi-head attention (Day 040) → GPT from scratch |
| ○ | Alignment & Beyond | LoRA, DPO · generative models · RL — after the block is trustworthy |
📂 Full derivations & code → AGI_Research — the notebook, not the summary.
Ship — what I build with
Mastered — rebuilt from axioms
Ship + Mastered live on the site → #stack · #services
CS50's Introduction to Programming with Python — 9 problem sets + final project · Cambridge, MA · Prof. David J. Malan · verify → · also on hariomlohardev.github.io#credentials
Available part-time for:
- Backend Systems — Django & FastAPI APIs, auth, DBs that stay up
- Mobile Apps — Flutter & Dart, one codebase → iOS & Android
- AI Features — LangChain + RAG over your own data, not a black box
- Automation & Scripts — scraping, pipelines, Python that reports clearly
→ hariomlohardev.github.io/#contact · or email hariomlohar.new@gmail.com (FormSubmit honeypot, _next=thanks.html) · clean code, honest timelines.
Featured — pinned below, live demos where it counts
More → Project archive · Blog & Daily Logs · RSS
"I don't just use the tools — I build them, so I actually understand how they work."
© Hariom Lohar — Lab Notebook No.01 · hariomlohardev.github.io · India · IST