David Okocha

19yo dev building Edgai • Compiling C from scratch to run local AI tutors on $30 phones with 0 internet • Engineering at Migranium • BYU-I.

[David Okocha @davidleonokocha]

I am David Okocha (@davidokocha086 ). I’m a 19-year-old engineer based in Lagos, and I’m breaking the current paradigm of AI development.

While 95% of AI startups are busy calling cloud APIs and building thin wrappers, I am building Edgai fully offline AI running on $30 Android phones. 🧵👇

[David Okocha @davidleonokocha]

@he_is_olabode dope dope! i'd check it out

[David Okocha @davidleonokocha]

I learnt something uncomfortable while building Edgai:

The model wasn’t the bottleneck.

Our curriculum architecture was.

I was building teaching around individual questions, when the real unit should have been the skill.

So i redesigned the layer underneath the AI.

[David Okocha @davidleonokocha]

@he_is_olabode love the energy @he_is_olabode

[David Okocha @davidleonokocha]

This changed how I think about AI tutors.

The hard part isn’t just getting a better model or writing better prompts.

It’s building curriculum infrastructure that knows:

what the student is learning,
what they misunderstood,
what can be revealed next,
and what the model is allowed to see.

Improve one concept once.

Every question that uses it gets better.

[David Okocha @davidleonokocha]

Now the architecture looks very different.

A question points to a skill.

That skill owns reusable teaching chunks.

Wrong answers map to misconceptions.

Hints move through an authored reveal ladder.

And the model only gets a prompt-safe view of what it is allowed to teach at that moment.

The LLM is no longer the curriculum.

[David Okocha @davidleonokocha]

@unicodeveloper Pedri just GPS’d that pass straight to Gordon’s foot like he paid for priority shipping. lmao

[David Okocha @davidleonokocha]

@unicodeveloper Flick got them running like rent’s due tomorrow 🐜 😂

[David Okocha @davidleonokocha]

While men slept he cooked.

Between this new ui looks better than what I had before.

I guess I am a bad ass designer 🧑‍🎨

[David Okocha @davidleonokocha]

still building

https://t.co/nmcP0ukewj

[David Okocha @davidleonokocha]

if you're new to programming you don't have to feel depressed. starting out, trust me it's very difficult and demotivating.

the first code i saw was html, when i laid eyes on it, i questioned the sanity of the person that created it 😂😂

be consistent even when it feels like nothing is entering with time it would eventually will and you would surprised! learn in bit and don't watch too many video tutorial build projects instead.

[David Okocha @davidleonokocha]

How I spent my Saturday

[David Okocha @davidleonokocha]

@passi0nfruitdsg @X hey @passi0nfruitdsg nice to meet you

[David Okocha @davidleonokocha]

Hey @X !

Looking to connect with investors 💰 and developers 🧑‍💻 in this space

- Offline Ai Inference on constrained devices
- African EdTech
- Ai Linux OS

[David Okocha @davidleonokocha]

Most AI demos stop at “the model runs.”

Mine starts there.

Edgai is pushing a 491MB model through real WAEC lessons on a 2GB Android phone at ~8–9 tok/s with no cloud inference.

The hard part isn’t chat. It’s memory, safety and correct teaching.

[David Okocha @davidleonokocha]

@thomasunise @X hi @thomasunise nice to meet you

[David Okocha @davidleonokocha]

Spent today testing whether a tiny AI model can actually tutor WAEC students offline on a 2GB Android phone.

Not “does it run?” does it avoid leaking answers, understand Pidgin, survive 30 minutes without cooking the phone, and still explain things correctly?

Building AI for real devices will humble you 😂

[David Okocha @davidleonokocha]

@manuelscript let’s work this

[OLUWADAMILARE @dam_ilare025]

Any single mistake like this, you gats start again..

[David Okocha @davidleonokocha]

the memory breakdown is where engineers get confused. model cards quote "0.4gb ram for bitnet 2b".

that is only the ternary non-embedding weights.

the 128k vocabulary embedding table is kept at higher precision (~0.5gb). add llama.cpp's kv cache, and real resident ram cost sits at ~1.1gb to 1.2gb.

so it doesn't run in 400mb, BUT it gives you a full 2B parameter brain inside the 1.1gb headroom of a 2gb phone!

[David Okocha @davidleonokocha]

standard llms spend all their compute cycles doing heavy floating-point matrix multiplications.

bitnet constrains every non-embedding weight strictly to {-1, 0, +1}.

that changes the math on an arm cpu.

you stop multiplying:
• +1: add input
• -1: subtract input
• 0: skip cycle completely

no floating-point mac units running hot just pure integer additions.

gsm8k math reasoning benchmarks show bitnet 2b hitting 58.38 vs gemma 1b's 31.16. that gap is massive when a student is revising math questions offline.

[David Okocha @davidleonokocha]

https://t.co/sSKGuJRCgG