A proposal for affordable AI that works offline, in Spanish, on equipment owned by the community: what we want to build, what it costs, how we will measure it, and how it can be sustained after the first grant ends. We are starting from zero, and we say so.
Executive summary
No technology is installed yet. We have legal status, community allies in La Cacica, a technical partner and a design. The first step is a low-cost seed pilot that generates our own evidence.
Open models small enough to run on a local computer or an edge device, solar-powered where needed, usable by voice in Spanish and working without internet.
The highest risks are our lack of a track record, technical knowledge concentrated in few people, and operating costs after initial funding. Every budget line answers one of them.
Seed pilot ≈ US$580 in equipment. Floor tranche ≈ US$7,200. Each later tranche is unlocked only if the previous one meets its indicators.
Part 1 — Starting point
A funder's first question is whether the team can deliver. Today CIDERT has no technology operating in the territory: an earlier attempt to fund a pilot in La Cacica did not succeed. This proposal starts from that fact and turns it into a strategy.
Legal status as a non-profit (NIT 902085017-1); community allies in La Cacica — the district's Community Action Board (JAC) and ACALCA, a peasant association being consolidated as part of the pilot; a strategic partner (Lógica Urbana S.A.S.); and a technical design for four lines of work.
Equipment, field data of our own, or trained models in operation. No figure on this page is a CIDERT result: they are external references or targets to be validated.
First, a low-cost seed pilot that produces real data within a few months. With that evidence and a university partnership, we apply for larger tranches.
It is easier to fund a team that shows one sensor working and real data than a team that shows a perfect design and has not started.
Part 2 — Baseline
In rural Colombia, fewer than half of households had internet in 2024, and most rural connections are mobile. A cloud-dependent service would leave most peasant families out. That is why the AI must run inside the territory, on equipment the community controls.
Colombia's AI policy of 14 February 2025: 106 actions to 2030 and about COP 479 billion, including AI adoption, talent and digital inclusion. [3]
The Ministry of Science opened seven calls for COP 2.7 trillion in 2025–2026, including one on biodiversity and productivity. [4]
Any model handling voices, photos of people or land data must comply by design.
Part 3 — Four lines of models
Each line has a local problem, a technique, its users and the metric that decides whether it works.
Coffee and cocoa pests are detected late, springs have no continuous records, and forest cover cannot be demonstrated to carbon markets.
Diagnostic accuracy against an agronomist, days of early warning, springs with continuous records.
Technical advice, procedures and funding calls reach rural districts late or not at all, and connectivity rules out cloud assistants.
Share of correct, sourced answers on a bank of real questions, and transcription error with the local accent.
Planting, harvest and sales decisions are made without forecasts.
Forecast error (MAPE) versus observed values.
Heritage is not catalogued digitally, visitors find no information offline, and community boards process documents by hand.
Items catalogued, visitor queries, administrative hours saved.
Part 4 — Risks
The six highest of nine identified risks. Score = probability × impact, each on a 1–5 scale.
| # | Risk | Level | Response |
|---|---|---|---|
| R1 | No track record. No technology operating yet, and a previous funding attempt failed. | High · 20 | Low-cost seed pilot, university partnership, tranche-based requests. |
| R2 | Knowledge concentrated in few people. | High · 20 | University as expert ally, training of local technicians, mandatory documentation. |
| R3 | No money after the grant ends for energy, replacement or staff. | High · 20 | Operating costs budgeted from day one; service income tested as hypotheses. |
| R4 | Hardware and exchange-rate volatility. | High · 16 | Edge devices instead of desktop GPUs; quotes in COP with validity dates. |
| R5 | Wrong answers from the assistant taken as true. | High · 16 | Verified documents only, cited sources, referral to a person on health or legal topics, evaluation before public use. |
| R6 | Personal and land data exposed or used without consent. | High · 15 | Data stays on local servers; informed consent; community data committee. |
Also identified: intermittent energy and connectivity (R7), no local training data yet (R8), low community adoption (R9). The scoring is preliminary.
Part 5 — Equipment and budget in tranches
Dollar prices are manufacturer or distributor list prices and exclude shipping, duties and import VAT. COP figures use the 30 September 2026 exchange rate (COP 3,341.23 per US$) [7]. Staff time is not yet costed.
| Item | Use | Qty | Unit | Total |
|---|---|---|---|---|
| Outdoor LoRaWAN gateway Dragino DLOS8N | Coverage for pilot sensors | 1 | US$315 | US$315 |
| Ultrasonic level sensor Dragino LDDS75 | First water-level series at 3 springs | 3 | US$88 | US$264 |
| Language assistant prototype | Small open model (≈ 4B parameters) on an existing 8–16 GB computer, tested with real questions | — | Existing | US$0 |
| Leaf photo collection | First labelled pest dataset with farmers' phones and agronomic validation | — | Existing | US$0 |
| Seed pilot equipment | ≈ US$579 · COP 1.93 M | |||
Seed pilot deliverable: a short report with real data (water-level series, labelled photos, assistant accuracy on a question bank), attached to every later request.
| Tranche | Main items | Priced items |
|---|---|---|
| Floor · full pilot + server in town | NVIDIA Jetson AGX Orin 64 GB (US$3,499), 3 gateways, 12 level sensors, 1.5 kWp off-grid solar kit | ≈ US$7,215 · COP 24.1 M |
| Base · municipal coverage (additional) | Second Jetson, 3 more gateways, 20 soil sensors, 2 solar kits | ≈ US$10,800 · COP 36.1 M |
| Expansion · regional replication | Only if earlier tranches meet their indicators | Not estimated |
Still to be quoted: rugged tablets, turbidity sensors, weather station, import costs and the technical team. Detailed tables are in the Spanish version.
Part 6 — Measurement and sustainability
A model counts when someone uses it and the decision taken with it can be verified. Almost every baseline is still "to be measured": measuring it is the first deliverable.
Local servers, informed consent, anonymized outputs, and a community data committee that decides on uses and access.
Monitoring services for the municipality and rural water boards, MRV evidence for carbon and PES projects, contributions from producer associations, and training with a university partner.
A rural AI project does not fail when the model makes a mistake. It fails when nobody in the territory knows how to fix it and there is no budget to do so.
Part 7 — Roadmap
Each phase starts only if the previous one meets its indicators. Months count from the moment seed funding is secured.
Seed pilot installed in La Cacica and first real data. Connectivity and energy measured by district, a bank of real questions, data committee, consent forms, university partnership.
Assistant server in town, 12 springs monitored, assistant evaluated before public use, first local technicians in training.
Forecasting models in use, wider LoRaWAN coverage, heritage catalogue, first service contracts.
Transfer to other peasant territories (TECAM), annual MRV report, operation funded mainly by service income.
Part 8 — What we are looking for
About US$580 in equipment plus technical and agronomic time for six months, to produce the first real data.
Advice on programs, calls and teams working on small AI, rural development and digital inclusion in Colombia and Latin America.
Universities and practitioners with experience in offline AI, Spanish speech models or community water monitoring.
Sources