Africa Deep Tech Challenge 2026 · Agriculture
The agronomist that never guesses.
Shuka answers maize, cassava, rice and tomato questions from real extension manuals — cited by page, entirely offline, on an ordinary 8 GB laptop. Where it has no source, it says so.
The problem
Good advice runs out before the road does.
Nigeria fields roughly one agricultural extension worker for every few thousand farmers — against a recommended one per thousand. The advice gap decides yields, and it is widest exactly where farming happens: rural areas with weak or unaffordable connectivity, on low-end hardware.
The obvious modern fallback — ask an AI — assumes bandwidth, cloud subscriptions and payment rails most smallholders don't have. So Shuka is built for the people farmers already ask: extension workers, agro-dealers and cooperative offices. One offline laptop at a cooperative becomes an answer desk for a whole community — and over the cooperative's own Wi-Fi hotspot, phones in the room can use it too.
Why grounding is the product
Small language models are confident liars.
These are real answers the same 1-billion-parameter model gave when we let it speak unsupervised. Each one could cost a farmer a season.
Grow cassava from seeds and harvest in about 3 months
Cassava is grown from stem cuttings and needs 8–24 months. A farmer following this loses the planting season outright.
DangerousPlant 30 maize seeds per hole
Recommended practice is 2 per station. Thirty seeds per hole destroys the stand and wastes a season's seed money.
DangerousStriga is a fungal disease — apply fungicides
Striga is a parasitic weed. Fungicide spend is wasted and the infestation keeps spreading.
DangerousHow it works
Every answer walks the same four steps.
STEP 01
Ask
A farmer's question, in plain language — diagnosis, planting calendars, input dosages.
STEP 02
Retrieve
On-device search over 1,308 passages from 8 license-verified manuals — FAO, IITA, CABI, IRRI. Milliseconds, no network.
STEP 03
Answer, cited
The model writes only from the retrieved passages. Every claim carries its manual and page number, like a printed bulletin.
STEP 04
Or refuse
No passage clears the relevance floor? The model is never invoked. Shuka says so and points to the extension office.
Not a chatbot
Built for the answer desk, not the chat window.
A cooperative's laptop serves a queue of people on market day. Shuka's features come from that queue.
Paper, on purpose
Printed answer sheets
One click formats the answer as a handout — question, advice, citations, date. The farmer leaves with something they keep, in the medium advice has always travelled here.
Institutional memory
The ledger
Every answered question is remembered on the laptop. A repeat — even reworded — is served from the ledger in under half a second instead of a fresh generation. Shuka gets faster the more the community asks it, and the record never leaves the machine.
Advice that travels
Copy as text
One tap turns the answer into plain text with its citations, ready to forward by WhatsApp or SMS — because that is how advice actually moves from the desk to the field.
The alternatives, honestly
What the answer desk has today, and what Shuka changes.
| At the desk | Printed manuals alone | Cloud AI chatbot | Shuka |
|---|---|---|---|
| Works without internet | Yes | No | Yes |
| Cost per question | Free | Data + subscription | Free after install |
| Finds the right page in seconds | No — browse 850 pages | Yes | Yes |
| Names its manual and page | It is the page | Rarely | Every answer |
| Refuses instead of guessing | — | Rarely | By design |
| Breadth beyond four crops | Whatever is on the shelf | Very broad | Not yet |
Shuka doesn't replace the manuals — it is the fastest honest index into them. And where a cloud chatbot is genuinely stronger today (breadth, languages), the comparison only applies where there's bandwidth and money to reach it.