Case study · Agricultural decision support · 2026
Krishyak
Better decisions begin before the harvest. An agricultural decision platform for India's small and marginal farmers, and the institutions that serve them.
- Role
- Product, ML, backend, frontend, cloud
- Status
- Public site and demonstration farm live · pilot-ready · farm accounts opening
The farmer's problem
Small farms carry large decisions alone.
- operational holdings in India
- 146Moperational holdings in India
- are small or marginal, under 2 hectares
- 86%are small or marginal, under 2 hectares
Agriculture Census figures as cited in the Krishyak briefing.
| Decision | Where farmers often turn | What it costs them |
|---|---|---|
| What to sow, how much to spend | Input dealer, last season's habit | Spend fixed before risk is understood |
| Is this leaf diseased? | Neighbours, shopkeepers, helplines | Late or generic diagnosis, wrong spray |
| Rain, heat and pest windows | TV, generic weather apps | Forecasts never become field actions |
| When and where to sell | Local trader, word of mouth | Little view of mandi prices or MSP |
| Which scheme applies | Portals, CSC operators, agents | Eligibility and documents unclear |
The farmer is the only place these answers meet. Krishyak puts them in one record.
The platform
Six steps, one loop: from what a farmer sees to what a farmer does.
- 01
Observe
Field boundary, crop cycle, notes, photos and soil readings.
- 02
Diagnose
A leaf photo becomes a likely cause and first steps, with abstention when unsure.
- 03
Analyse
Weather, satellite indices, mandi prices and MSP, each with source and date.
- 04
Simulate
Monte Carlo season runs compare current, optimised and worst-case plans.
- 05
Recommend
Fertilizer schedules, inspection prompts and scheme guidance, with reasons.
- 06
Act
The farmer records what was done and what happened. The farmer decides.
Built for trust: every conclusion carries its source and date, uncertainty is shown rather than hidden, and the farmer stays the decision-maker. A satellite signal is a reason to inspect, not a diagnosis.
Capability atlas
Thirteen capabilities on one data spine.
- Verified live
- In public app
- Built
Verified live: tested end to end with real providers. In public app: shipped and tested in the application. Built: implemented and tested, awaiting a live feed, partnership or field trial. Status per the readiness audit, October 2026.
- Verified live
Field records
GPS boundaries or area, crop cycles and a dated timeline
- Verified live
Leaf photo check
EfficientNetV2B0, 38 classes; abstains below 0.75 confidence
- Verified live
Satellite indices
Sentinel-2 NDVI, NDMI at 10 m and NDRE at 20 m, with cloud masking
- Verified live
Weather context
48-hour and 7-day forecasts for a mapped field
- Verified live
Voice and language
Speech in and out via Sarvam, with written fallback
- Verified live
Privacy controls
Purpose consent, export and verified deletion
- In public app
Season simulator
Current, optimised and worst-case plans compared
- In public app
Farm economics
Cost of cultivation, revenue, profit and ROI
- In public app
Fertilizer and soil
Nutrient gaps, doses, splits and an organic mode
- In public app
Pest risk
Weather-driven risk scoring with spray-window rules
- Built
Market and MSP
Mandi adapter with dated price history and MSP reference
- Built
Scheme guidance
PM-KISAN, PMFBY and Soil Health Card checklists with official links
- Built
Offline and pilots
Field mode, consented cohorts and pilot reports
04 · Crop health and AI vision
A leaf photo becomes a likely cause and practical first steps.
EfficientNetV2B0, trained on hash-pinned public data with leaf-level splits, served on LiteRT with 247 of 247 mobile-model parity checks.
- — Abstains below 0.75 confidence or on a crop mismatch.
- — Gives no pesticide dose advice.
- — Always says who to confirm with before spraying.
- Validation accuracy
- 94.39%5,416 images · 95% interval 93.7–95.0%
- Held-out test accuracy
- 93.71%8,566 images · 95% interval 93.2–94.2%
- Test balanced accuracy
- 91.52%Same held-out set
- Test macro F1
- 0.915Same held-out set
Trained and evaluated on public datasets (PlantVillage and PlantDoc, CC BY 4.0). These are laboratory-style results, not field accuracy: the next release gate is Indian field photos with blinded expert labels.
14 supported crops
- Tomato
- Potato
- Maize
- Soybean
- Capsicum
- Grape
- Apple
- Orange
- Peach
- Cherry
- Strawberry
- Blueberry
- Raspberry
- Squash
Season simulator
See the season's range of outcomes before spending.
500 Monte Carlo runs per plan vary rainfall ±20%, pest pressure ±0.15, fertilizer ±15% and price ±10%. The example below is the illustrative 2-hectare run shown in the Krishyak briefing; it is a model, not a farm result.
Current plan
₹45,409
modelled profit · ROI 66%
- Yield
- 2,484 kg/ha
- Risk score
- 31 / 100
Optimised plan
₹68,588
modelled profit · ROI 91%
- Yield
- 3,131 kg/ha
- Risk score
- 27 / 100
Worst case
−₹60,170
modelled loss · ROI -67%
- Yield
- 819 kg/ha
- Risk score
- 59 / 100
Language and voice
22 scheduled Indian languages, plus English.
Every screen ships as a static, versioned language pack, cached for offline use. Urdu, Kashmiri and Sindhi run right to left; Manipuri uses Meitei Mayek and Santali uses Ol Chiki. Speech in and out via Sarvam, with consent and a written fallback. Native-speaker review of every pack is on the roadmap.
- Assamese
- Bengali
- Bodo
- Dogri
- Gujarati
- Hindi
- Kannada
- Kashmiri
- Konkani
- Maithili
- Malayalam
- Manipuri
- Marathi
- Nepali
- Odia
- Punjabi
- Sanskrit
- Santali
- Sindhi
- Tamil
- Telugu
- Urdu
- English
Architecture
Modular, privacy-first, built for patchy networks.
- FarmerInstallable PWA in Next.js and React, offline field mode, passkey sign-in
- APIFastAPI modular monolith, consent and deletion rights, durable job queue
- IntelligenceEfficientNetV2B0 on LiteRT, Monte Carlo simulator, fertilizer and pest engines
- DataPostgreSQL with PostGIS, Redis usage budgets, private photo storage
- ProvidersCopernicus Sentinel-2, Open-Meteo, Sarvam speech and translation, data.gov.in
Readiness audit: 295 backend tests, 405 frontend tests across 65 suites, CI on Chromium, Firefox and WebKit, 0 known runtime vulnerabilities at audit.
Where it stands
Pilot-ready. Not yet field-proven.
Built
Done
Tested
Done
Verified live
Done
Farmer pilot (next)
Next
Field benchmark
Ahead
Scale
Ahead
Recognition
The founder presented Krishyak to the Union Agriculture Minister and the Director General, ICAR, at the Viksit Bharat Young Leaders Dialogue, January 2026.
Not yet, and not claimed
No independent field validation, funded pilot, institutional partnership or revenue has happened yet. Proposed pilots and partnerships are plans, and Krishyak is not presented as a government-approved service.
Running a programme for farmers?
FPOs, NGOs, CSR teams and agriculture programmes can discuss a consented, carefully scoped Krishyak pilot. We will tell you plainly what the platform can and cannot do today.
Or book a free 30-minute call (opens in a new tab) with the founder.
Prefer email? hello@ameyor.in