Winner · Agriculture trackSmall AI for DevelopmentHackathon 2026
Cooperate through messaging.
Dure: the Korean tradition of mutual aid between farmers.
A good harvest is not a good price.
Noor grows coffee on the slopes of Letefoho, Timor-Leste. Is the market treating her well?
The buyers are mostly in the capital, Dili.
of Ermera's coffee is bought by just four traders in the capital (survey of 100 farmers, 2014).
Cristovão, Bogor Agricultural University, 2015What if her neighbours sold with her?
40 kg · Noor1,890 kg · 22 farms
Flooded. Only one trader's truck makes it.
One truck, booked for the whole lot.
- Traderon the road$2.25
- ExporterDili$2.81
- RoasterDili$2.79
- ExporterDili$2.76
- Café chainDili$2.74
$2.25$2.81per kg+25%
Same coffee. More of it, and buyers compete. Illustrative example.
This is cooperation. This is Dure.
of 239 studies found that farmer organisations raised their members' incomes.
Bizikova et al., Nature Food, 2020So why isn't she benefiting from the Letefoho coffee cooperative?
The catch
Running a cooperative in Timor-Leste is not easy.
- 3–4years
- Time. Recruiting, then paperwork.
- $1,000
- Capital. A coffee household earns about $250 a year.
- power.and then
- Power game. Women farmers are easily left out.
KOICA value-chain cooperative, Timor-Leste; Decree-Law No. 16/2004; The Irish Times, 2013.
So Noor still sells alone, at the trader's price.
What if the village kept the cooperative's advantages, but dropped the cooperative?
Introducing
Cooperate through messaging.
- Pooling the harvest
- Buyers bidding for the whole lot
- One shared truck and pickup point
- A record that earns trust
- A legal entity
- $1,000 in share capital
- A board to run, a leader to fight over
- Unequal voice: a few decide for everyone
What stays
- Pooling the harvest
- Buyers bidding for the whole lot
- One shared truck and pickup point
- A record that earns trust
What goes
- A legal entity
- $1,000 in share capital
- A board to run, a leader to fight over
- Unequal voice: a few decide for everyone
No internet. No app. Just SMS.
How it works
One week, by text message.
Noor's week, on her basic phone. Behind each message, Dure does a cooperative office's work with the buyers. Tap a step to open its messages.
1. A price and a plan to start the dayAt 6:30 Dure sends a fair price, built from past trades, weather and the road.
A $3.17–3.23 · B $2.80–2.86 · C $2.12–2.18
Selling? Reply in your own words.
Heavy rain, Gleno road flooded: only 4 buyers.
Ask at least $3.17; photo before 10:00.
2. From Tetum text to quality gradingNoor writes in her own words and sends one photo. Dure reads both.
Coffee · 40 kg · grade A · ask $2.84/kg
Now send one photo of your beans.
A person checks every sack at pickup.
3. Leverage with a blended basketEvery grade goes into one lot. Together, smallholders have bargaining power.
Bidding opens $2.74/kg, closes 17:00
4. Interactive fair auctionWhen the lot is full, buyers bid. Dure sets a reasonable opening price.
A $3.14 · B $2.78 · C $2.11
5. Collective supply chainDure books one truck. The grade is confirmed by hand at pickup.
Grade A confirmed by hand
6. Simple and credible paymentThe buyer prepays; after delivery each farmer gets her share.
7. Every deal builds trustTrades, punctuality and accurate quality build a reputation score.
Likely to deliver: 83% (was 79%)
Illustrative example — names, volumes and prices are not real data.
For the curious
Technical walkthroughTetum texts · photo grades · reputation · two markets
AI that understands Tetum.
A small model (Qwen2.5-0.5B) plus a rules engine, trained on about 3,700 Tetum and English texts. Dialects and spelling slips are fine.
Open-source Qwen, trained on about 3,700 texts incl. 250 real Tetum (Labadain-30k+, CC BY 4.0).
- kafe
- Cropcoffee
- 40kilu
- Quantity40 kg
- diak los..
- GradeA · preliminary
- fan 2.84
- Ask (hidden)$2.84 / kg
- bondia! bele? ok
- Intentoffer to sell
AI that knows coffee.
A small vision model trained on 1,200 bean photos grades coffee A, B or C, right 79% of the time. A hand check sets the final grade.
Criteria: SCA green coffee defect classification. Bean photos: Pre-Roast Coffee Bean Grading Dataset (MIT) and Wikimedia Commons.
1Find the beans done
2Look for defects
3Compare with graded photos
Some of our training photos · Wikimedia Commons: H. Ulver, M. C. Wright, F. Quijano (CC BY-SA 4.0); Forest & Kim Starr (CC BY 3.0)
AI that measures trust.
Trust comes from trades, punctuality and steady quality, not farm size. Floods and other force majeure never count.
In the demo, reputation is computed from pre-accumulated synthetic trade records.
- 14sales
- 13on time
- 92%grades confirmed
- 1dispute
- 1flood week, not counted
- 12sales
- 9on time
- 75%grades confirmed
- 3disputes
- 0flood weeks
Dure AI's estimate, ± a few points. Land and volume aren't counted.
AI that runs two markets.
The blended pool is an open market with scale. The fixed pool links part of trusted farmers' output to steady buyers.
In the demo, the fixed pool runs on pre-accumulated synthetic records.
Blended pool
every farmer sells here, every week
harvestreputation
Fixed pool
one grade · a fixed buyer · every week
higher reputation, more steady orders
Dure AI sets the split every week
How well does AI read Tetum texts?Tetum Benchmark Index
Large models overlook Tetum. A small model plus a rules engine gets Dure close to a large model.
Method. 60 synthetic SMS written by Claude Sonnet, frozen before scoring; every field must be right. On 64 newer texts never used to build Dure: 86% (rules engine alone 78%). Synthetic benchmark, not field results; the web demo reads offers with the rules engine only. Measured 23 Oct 2026.
An automated Registry and AI policy suggestions.From the slope to the ministry
A farmer Registry is costly to keep by hand. Every Dure deal leaves a record, so the Registry builds itself.
1 · Daily accumulated trades
| week | kg | grade | $/kg | flag |
|---|---|---|---|---|
| LTF-002 · Lebudu · 0.8 ha | ||||
| wk 3 | 60 kg | B | 2.49 | |
| wk 4 | 55 kg | B | 2.51 | |
| wk 5 | 50 kg | B | 2.47 | |
| wk 6 | 40 kg | C | 1.90 | mould |
| wk 7 | 35 kg | C | 1.92 | mould |
| LTF-009 · Ducurai · 2 ha | ||||
| wk 3 | 60 kg | A | 2.84 | |
| wk 4 | 65 kg | A | 2.88 | |
| wk 5 | 60 kg | A | 2.86 | |
| wk 6 | 70 kg | A | 2.90 | |
| wk 7 | 60 kg | A | 2.86 | |
+ 593 more trades · every SMS lands here the same day
2 · Data analysis
Quality signals by week, whole pool
Insect damageLow grade (C) shareBlack or mouldy beans
One of the Registry's charts · synthetic data
3 · AI brief
Mon 27 July · 06:00 · Letefoho pool
Lebudu: mould on 60% of photos after the rain
against 3% in the other four villages
- kg per offer down 35% since week 1
- grade C share 10% → 70%
- 7 of 10 offers sat out last week
Not a diagnosis. For expert review:an extension visit to Lebudu this week and a check on drying space
every number links to its rows in the Registry
AI is not a silver bullet. Noor knows farming better than Dure. Dure helps the extension officer make the most of two visits a year.
Open the sample Registry→Synthetic data for eight weeks in Letefoho.
What are the real field problems?Roads, records, power
The road washes out.
Dure books the truck and the pickup point. It can't fix the road, and doesn't pretend to.
Who sees the records?
Farmers' names and numbers stay with Dure. The ministry's Registry shows pseudonymous IDs and village totals, and only for farmers who said yes.
The power or the server goes down.
The rules engine keeps taking offers by text. The small model rechecks them later, and any change goes back to the farmer to confirm.
Electricity is a real problem in Timor-Leste.
- 40%
- of firms had power cuts in a year, losing 9.4% of sales.World Bank Enterprise Survey, 2015
- 24 h
- outages in Viqueque. Dili's hospital runs on generators each time.Tatoli, 2022–23
- 36%
- of Dili customers never report an outage.TANE survey, 2022
Cooperate through messaging.
Try the demo→
Founder
Tae Yoon Moon
University of Seoul · International Relations & Economics
Thanks for paying attention, and enjoy the demo!