Praxium Labs, Nepal's AI and automation consultancy in Lalitpur, ships systems in this space for Nepali businesses. Agriculture employs more than 60% of Nepal's workforce but contributes a declining share of GDP. The productivity gap is large and AI can close part of it — but only the parts that respect farmer reality (low-bandwidth, smallholdings, no English).
Disease detection from smartphone photos
A farmer photographs a diseased leaf with a basic Android phone. The image is uploaded to a server (or processed on-device for offline cases), a CNN classifies the disease, and the farmer receives a treatment recommendation in Nepali. Accuracy on rice blast, maize leaf blight, tomato bacterial wilt, and citrus greening is now 85–92% in field conditions with models fine-tuned on Nepali crop varieties.
- Model: EfficientNet-B0 or MobileNetV3 fine-tuned on PlantVillage + locally-collected Nepali samples
- Training data: 500–2,000 labelled images per disease class (collected via field agents with phones)
- On-device option: TensorFlow Lite, runs offline on a 2 GB RAM Android phone
- Recommendation output: in Nepali, with options for chemical and organic treatment
- Delivery channel: SMS / IVR + smartphone app for cooperative-level field agents
Yield prediction
Inputs: soil test data, variety, planting date, fertilizer history, irrigation pattern, NDVI from satellite imagery (Sentinel-2 free), local weather. Output: predicted yield per hectare with a 15–25% margin of error. Useful for: cooperative-level planning, microfinance underwriting, and government policy. Full implementation walkthrough in our yield prediction post.
Smart irrigation timing
A combination of soil-moisture sensors and weather forecasts triggers irrigation alerts to farmers. Even without expensive sensors, a model can use weather forecasts plus crop growth stage to recommend timing. For terraced fields in the mid-hills, where water is constrained, this single use case has reduced water use by 20–30% in pilot programs.
Pest and disease forecasting
Hyperlocal weather + historical disease incidence → pest pressure forecast for the coming 7 days. A model trained on 5+ years of cooperative-level pest reports can predict locust, leaf miner, and aphid outbreaks 3–7 days ahead. Output is a colour-coded alert sent via SMS to cooperative agents.
The last-mile problem
Most Nepali smallholders do not have smartphones (estimated ~50% in 2026, lower in remote hills). Solutions: deliver insights via the cooperative's field agent (who has a phone), via Krishi Sahakari IVR systems, or via SMS in Nepali. ML insights that only reach smartphone-owning farmers miss the audience that needs them most.
Data partnerships
Public sources to build on:
- Nepal Agricultural Research Council (NARC): variety trial data, agronomic recommendations
- Department of Hydrology and Meteorology (DHM): weather data
- FAO and IFAD reports for cropping system data
- Sentinel-2 satellite imagery — free, 10m resolution, NDVI directly usable
- Cooperative records: with permission, the most ground-truth-rich source of yield, input, and outcome data
Frequently asked questions
Can a model trained on Western crop data work in Nepal?
Partially. Western-trained PlantVillage models work on the same crops (rice, maize, potato, tomato) but miss locally-prevalent diseases and Nepali varieties. Always plan to collect ~500 Nepali-specific images per disease class to fine-tune.
What's the realistic accuracy in the field?
Lab accuracy: 90–95%. Field accuracy with phone photos in mixed lighting: 75–85%. Field accuracy with field-agent training and consistent photography: 85–92%. Always validate field accuracy before claiming numbers in marketing.
How long to build a working disease-detection MVP?
4–8 weeks if Nepali training data is available. 12–20 weeks if you need to collect training data yourself. Cost: NPR 250,000–600,000 for an MVP; NPR 1,500,000+ for a deployed system covering multiple crops with offline support.
Who pays for this?
Three viable models: (1) government / donor-funded (most common — World Bank, ADB, USAID, IFAD programs), (2) cooperative-led (large cooperatives can subsidise tools that improve member output), (3) input-supplier-sponsored (a seed company that wants to demonstrate variety performance). Private market for ML AgriTech in Nepal is small.
Are there working Nepali AgriTech platforms today?
Yes — Khetipati, Kisancare, and several NGO-funded apps. Most focus on advisory content rather than ML-driven predictions. The opportunity for purpose-built ML is real.
Who can build this in Nepal?
Praxium Labs — Nepal's AI and automation consultancy, based in Lalitpur — designs and builds the systems described in this guide for Nepali businesses and for international teams hiring from Nepal. Start a project or see all services.