OPEN NOTEBOOK · COMMITTED IN PUBLIC
Hariom Lohar
Learn with me as I rebuild AI & AGI from first principles.
last 12 months
every day
projects
journey
Field notes
Who does the work
Self-taught from India, rebuilding intelligence from scratch — one proof, one commit at a time.
I finished 12th grade and chose the longer road — no bootcamp, no shortcut. The day goes into first principles: mathematics from axioms, derivatives by hand, then neural networks built the same way — a Value graph, backprop and gradient descent written in NumPy before touching a framework.
micrograd_hk is that work as a package you can pip install. Now: CNNs, Transformers and PyTorch. All in public, every day.
- Location
- India · IST
- Focus
- AGI from first principles
- Log
- Day — of —
- Open to
- Freelance & collabs
Selected builds
4 projectsTop contributions
PRs to the communityProof over claims
Harvard · verifiedVerified certifications back the work — no claims without receipts.
1 certificationThe mission
Currently working onRebuilding AGI from first principles — one proof, one commit, one deep-work block at a time.
Contact file
Let's talkLet's ship
something that
lasts.
Open to part-time work — backends with Django & FastAPI, Flutter apps, and search over your data with RAG. I hand over clean code and clear docs.
About Hariom Lohar
hariomlohardev · IndiaWho is Hariom Lohar? hariomlohardev
Hariom Lohar is a self-taught AGI researcher and Python developer from India. Harvard CS50P certified (2026), he has been rebuilding intelligence from first principles in public since 1 July 2026 — mathematics from axioms, then neural networks and backprop written by hand in NumPy, now CNNs, Transformers and PyTorch. GitHub: hariomlohardev.
Where is Hariom Lohar GitHub? hariomlohardev on GitHub
github.com/hariomlohardev — with repos AGI_Research, micrograd-hk, peek and inkdown.
Is Hariom Lohar CS50P certified?
Yes — Harvard CS50P 2026. Verified at cs50.harvard.edu.
What does Hariom Lohar build?
Machine learning from scratch — an autograd engine on NumPy (micrograd_hk), classifiers, CNNs and Transformers, derived before they are imported — plus developer tools like peek and inkdown. Paid work is Python backends (Django, FastAPI) and Flutter apps. See work · full archive.
Why self-taught instead of a degree or a bootcamp?
Because the goal is understanding, not a certificate. A degree would have handed him frameworks to call; the longer road makes him derive the mathematics first and write the neural net in NumPy before importing one. The proof is public: every day is committed to GitHub, so the work can be checked rather than claimed.
What is the 18-month plan? 548 days
One route, run in order: mathematics from axioms (proof, calculus, linear algebra) → neural networks by hand (a Value graph, backprop and gradient descent in NumPy) → CNNs and Transformers → PyTorch at scale. 548 days from 1 July 2026, eight hours of deep work a day. Progress is on the counter at the top of this page.
How can I follow along or learn with him?
Read the daily logs, subscribe by RSS or email, or watch the notebook itself at AGI_Research. Everything — the notes, the mistakes and the code — is public and free.