Downshift · Pre-Seed · 2026
A devoted tutor for every child on Earth.
Starting where there are none.
An adaptive AI tutor that teaches children to read, write, and do arithmetic. Fully offline, on the phone their family already owns.
The problem
In the Andes and the Amazon, the teacher exists on paper. The child is on her own.
Why now
The $15M Global Learning XPRIZE put offline tablets in front of 2,700 unschooled Tanzanian children. In 15 months, non-readers were cut in half. No teacher present.
The winners hand-authored 20,000+ activities over years. A frontier model now writes an entire scaffolded curriculum in batch. The expensive part of 2019 costs almost nothing in 2026.
Android holds 90 to 96% of every target market. No hardware to ship, no lab to build, no OLPC. The classroom is already in the family's pocket.
Proof the model works. Content at zero marginal cost. Hardware already deployed. Nobody has combined the three.
The product
Text + drawings · the complete tutor, works anywhere
+ Songs · when a connection passes by
+ Video · when there is real bandwidth
+ Live dialogue · when the village gets its antenna
Same tutor, same curriculum, same child. The product meets the household where it is, and gets richer as connectivity arrives.
How it works
A mastery engine decides what to teach next. Interpretable, sub-100ms, offline. No hallucination surface.
Content is authored in the cloud, reviewed by humans, and shipped as signed packs. The device plays; it never improvises facts.
Recognition and a warm voice run locally. The child speaks, draws, taps, and traces. No data leaves the phone.
Signed packs pass by Bluetooth, Wi-Fi Direct, or SD card. One connected phone bootstraps an entire village.
The generative AI does the authoring, once, where we can review it. The child gets a system that is correct by construction.
Unit economics
Every connected competitor pays per token, per child, forever. We pay once, at authoring time.
Market
One Spanish curriculum serves ~19 countries. Indigenous packs (Quechua, Aymara, Guaraní, Maya) layer per region. Visuals are language-neutral; narration is a text swap.
Latin America invented teaching remote children without teachers: Radio Sutatenza reached 900+ municipalities by radio and workbooks. We are its smartphone-era successor.
Business model
An adoption price, deliberately at the floor. Low enough to never gate a child, real enough to signal value.
$15-30/month · the diaspora. A relative abroad sponsors a child back home. Remittances to LMICs reached $656B a year. The monthly progress report is the receipt that makes a subscription feel like school fees.
Mining and energy companies fund rural schools near operations as social license. Antamina alone put S/14M into two rural school programs and has run one with Enseña Perú since 2017.
One CSR deal funds thousands of children plus the school's satellite antenna for sync. Partnerships buy the beachhead; the subscription scales north.
Competition
The tutor. Everyone built content or platforms. Nobody built someone the child returns to.
We searched for a fully offline, permissionless, parent-downloaded child tutor in Spanish and Portuguese and demonstrated its absence. The niche is open, and it is contested only by companies structurally unable to enter it.
Defensibility
Our bar: delayed, unassisted, independently administered post-tests. The whole category fails it. Proven efficacy is the one asset money can't copy quickly.
Songs and characters are the only free acquisition ever proven at scale in these markets. Ubongo reaches 11M viewers a week on that engine.
A per-child record of mastery, compounding for years. It powers the tutor, the parent report, and the sponsor receipt.
Scoring a child reading aloud, on-device, in Spanish and Quechua. No open solution exists. We build it where the speech data actually exists.
We do not compete on running models offline. Google gives that away. We compete on the tutor, the evidence, and the character.
Why us
Everything the Primer needs, we already operate: explainer generation, music-based learning, voice cloning, an eval culture that refuses unmeasured claims, and years of practice making software feel like someone rather than something.
The plan
Arithmetic tutor MVP in Spanish + Quechua, running offline on a $70 Android. Learner model live. Diaspora price test running before code is even finished.
First funded deployment conversation: Enseña Perú and a mining CSR program. Pilot design for 2,000 children. Reading assessment prototyped on existing Spanish child-speech corpora.
Pilot in the field. Independent efficacy study designed to the bar the category fails. Second country on the same spine. The evidence becomes the moat.
Arithmetic needs no speech recognition, so it ships first. Reading follows where the corpora already exist. Every risk has a sequenced answer.
The ask
In range for the round: LatAm AI-education pre-seeds cleared at $1.4-1.6M this cycle. Nobody in the cohort is rural, offline, or evidence-gated. That position is open, and it is ours to take.
The Primer
The phone is already in their pocket.
Manuel Zamora · Downshift
Rural Peru → all of Latin America → every child on Earth