AniAI: A crop doctor in every farmer's pocket, speaking their dialect

Problem Statement

Agriculture still employs about a quarter of Filipino workers, yet farmers and fisherfolk remain among the poorest sectors in the country. Pests and diseases, including rice tungro, bacterial leaf blight, fall armyworm in corn, Fusarium wilt (Panama disease) in Cavendish banana, and the coconut scale insect (cocolisap), regularly wipe out a large share of harvests. Each Department of Agriculture technician covers many barangays, so a farmer can wait days or weeks before an expert looks at a crop. In that time the infection spreads, or the farmer buys the wrong pesticide. When the advice does come, it is usually in English or Tagalog and ignores the incoming rain or typhoon that will wash the treatment away.

Proposed Solution

AniAI is an offline-first mobile app with an SMS/voice fallback. The farmer photographs a leaf, stem, or fruit. A small vision model running on the phone, trained on rice, corn, banana, and coconut disease and pest images, returns the likely diagnosis, a confidence score, and how severe it looks. Once the phone gets even a 2G/3G connection, the case syncs to the cloud. There, an LLM grounded in PhilRice, DA-BPI, and PCA guidance writes a treatment plan in the farmer’s dialect and adjusts it to the PAGASA forecast, for example “Do not spray today, heavy rain is expected within 12 hours; spray Thursday morning instead.” Farmers with basic feature phones can text a keyword or call an IVR line and hear the advice. Each diagnosis is geotagged, so DA and LGU agriculturists get a live outbreak heatmap and early warning of where to send technicians.

Target Market

• Primary users: Smallholder rice, corn, banana, and coconut farmers, roughly several million households across Luzon, Visayas, and Mindanao, many of them in areas with patchy connectivity.
• B2G buyers: DA regional field offices, PhilRice, PCA, the Bureau of Plant Industry, and provincial and municipal agriculture offices, which can use it for outbreak surveillance and as an extension tool.
• B2B buyers: Farmer cooperatives, agri-input retailers, rural banks, and crop insurers such as PCIC, which can use it for claims evidence and risk scoring.
• Size: Nationwide reach through DA’s extension network. Over time it can cover ASEAN rice-growing markets.

How AI Is Used

• On-device computer vision (MobileNet/EfficientNet-Lite, TFLite) classifies diseases and pests offline. Without it, a diagnosis needs a human expert who is days away.
• LLM with retrieval (RAG) over official crop protection manuals turns a bare label into a treatment plan the farmer can act on and that complies with local rules. Without it, the farmer gets a disease name and no next step.
• Machine translation and speech (TTS/ASR for Filipino languages) deliver the advice in dialect and by voice for farmers with low literacy. Without them, adoption collapses outside Tagalog and English speakers.
• Weather-aware reasoning combines PAGASA forecast data with the treatment schedule so farmers do not waste chemicals or spray right before a typhoon.
• Spatial anomaly detection on clusters of geotagged reports flags emerging outbreaks for DA before they spread across a province.

PLDT / Smart / ePLDT / Rokid Angle

• Smart SMS/IVR and USSD carry the feature-phone channel and generate recurring B2G messaging revenue. Zero-rated access can be bundled into Smart prepaid promos.
• ePLDT cloud and VITRO data centers host the models, RAG store, and outbreak dashboard. Keeping farmer data in the country supports DPA compliance.
• PLDT Enterprise B2G sales can package AniAI with connectivity for DA, LGUs, and cooperatives.
• Rokid AR glasses let DA technicians do hands-free field scouting, with diagnoses overlaid on the plants as they walk the paddies.
• Smart network data helps plan where coverage gaps hurt farmers the most, which links to rural connectivity rollout.

Demo Scenario (3 minutes for judges)

1. Mang Ramon, a rice farmer in Iloilo, notices yellow-orange leaves. On stage, a teammate photographs a real or printed infected leaf with airplane mode turned on.
2. The app shows “Likely rice tungro, 87% confidence, moderate severity” with no internet connection.
3. Airplane mode is turned off. The case syncs, and a Hiligaynon voice message plays: remove infected hills, control green leafhoppers, and wait for the dry window because PAGASA expects rain tonight.
4. A second teammate texts “ANI MAIS” from a basic phone and receives the fall armyworm advice by SMS.
5. The DA dashboard lights up with a cluster of tungro reports in three Iloilo barangays and sends an automatic alert to the municipal agriculturist.
6. Close: “Before AniAI, Mang Ramon waited two weeks. Now he waits two seconds.”

30-Hour Build Plan

Suggested stack: TensorFlow/Keras → TFLite, React Native or a PWA, FastAPI, Gemini or open LLM (e.g., Llama/SEA-LION) with a vector DB, Google Cloud TTS or open TTS, PostgreSQL/PostGIS, Leaflet. Public data / APIs: PlantVillage and rice leaf disease datasets (Kaggle/Hugging Face), PhilRice Rice Doctor and DA-BPI pest guides, PCA advisories, PAGASA forecasts, Open-Meteo API, PSA OpenSTAT agriculture statistics, OpenStreetMap barangay boundaries.

Risks & Privacy Notes

• Misdiagnosis risk: Show confidence scores, always recommend confirmation by an agriculturist when confidence is low, and never recommend banned pesticides. The RAG sources are limited to official guidance.
• Dataset bias: Public datasets are mostly from other countries. Add augmentation and plan for a locally collected dataset with PhilRice/state universities.
• Low connectivity: Inference runs on the phone first, and SMS is the fallback.
• Data Privacy Act of 2012: Farmer names and phone numbers are collected only with consent, location is coarsened to barangay level on public dashboards, and data is encrypted and stored locally.
• Clause 24 compliance: No personal or confidential data is entered into third-party AI tools during development. AI use is disclosed, and dataset and model licenses (e.g., CC BY) are respected and credited.

AI Use Disclosure (draft)

We use TensorFlow/MobileNet (open source) for image classification and Gemini/an open LLM for retrieval-grounded advice and dialect translation. We also use a TTS model for voice output. ChatGPT/Gemini and GitHub Copilot helped draft boilerplate code and pitch copy. All diagnoses, treatment text, and code were reviewed and tested by team members, and agronomic content was checked against DA/PhilRice references.

Rubric Scoring