See crop risk early,
apply the right biological

Live satellite, weather and soil become an explainable risk score, a matched biological product and a money case, every season.

Sasya EngineTM

Sasya unites satellite, weather and soil signals and matches the right biological product to your field, with the economics to back the call.

How it works 2:00
Leaf disease scan EfficientNet-B4
Diseased leaf with dark fungal lesions used to demonstrate the scanner
Tomato
Septoria leaf spot
96%
confidence
fungal blight Grad-CAM ready
Risk MODERATE
NDVI 0.61
Built for HCHACK CORE 2026 AANNAM.AI, IIT Ropar SSyngenta Biologicals GGoogle

"A gap exists between the potential of biologicals and what farmers can actually access."

01

Thin efficacy data

Little proof of what works across climates, soils and cropping systems.

02

No single view

Nowhere to see field risk and match it to the right biological in one place.

03

No season tracking

Recommendations never learn from what happened in the field last time.

04

Fragmented access

Growers, distributors and retailers work off different, disconnected signals.

The closed loop

Visualise, recommend, then learn from the field

Three stages run as one loop on real data. Each field you analyse feeds the next recommendation.

Stage 01 / Risk intelligence

Read the field before symptoms show

Sentinel-2 NDVI and NDMI, an Open-Meteo forecast and SoilGrids soil combine into one explainable 0 to 100 risk score with early warnings.

Vegetation, weather and soil sub-scores
Satellite NDVI tile overlay on the map
Detected problems handed to Stage 02
live signals
64
composite risk
Stage 02 / Biological advisor

Match risk to the right product

A hybrid rule and weighted scorer filters the biological catalog on crop, soil, climate and detected problems, then ranks matches with confidence and a plain-language reason.

What to apply, when, and why
Transparent feature contributions
Grounded advisory in seven languages
ranked
Trichoderma biofungicide0.88
Pseudomonas biocontrol0.74
Seaweed biostimulant0.61
Stage 03 / Continuous learning

Every outcome retrains efficacy

Farmers log what actually happened. Per-product efficacy updates, the model version bumps, and Stage 02 recommendations shift with the evidence.

Outcomes persisted per field and season
Before and after efficacy lift shown
The loop closes, then repeats next season
retrained
Trichoderma0.750.80+0.05
Azospirillum0.700.66-0.04
Silicon biostim0.680.64-0.04

One platform, the whole season

Risk, recommendation, disease vision, advisory and economics, all explainable and all wired to the same backend.

Risk intelligence

Sentinel-2 NDVI and NDMI, live weather and SoilGrids soil become one composite risk with early warnings and a satellite map overlay.

Track 1
Close-up of a leaf with yellow chlorosis and necrotic holes

Disease vision with Grad-CAM

Scan a leaf photo, get a class in 38, and see exactly where the model looked.

Biological advisor

Ranked products with confidence and reasons.

Yield and ROI

Transparent attribution on real MSP prices.

Multilingual chat

Grounded advice in Hindi, Punjabi, Tamil and more.

EfficientNet-B4 held-out test · 10,878 images
99.1%test accuracy
0.988macro-F1
0.991weighted-F1
38disease classes
17.8 msper scan
Per-class F1 across 38 classes
F1 ≥ 0.99
26
0.95 – 0.99
10
below 0.95
2
trained on PlantVillage · 14 cropsGrad-CAM explainable

The model shows its work

EfficientNet-B4, trained on PlantVillage to 99% accuracy across 38 classes, classifies the leaf and overlays a Grad-CAM heatmap so an agronomist can trust the call, not just take it.

1

Upload or drop a leaf photo

Any crop in the catalog, straight from a phone camera.

2

Classify and explain

Top predictions with probabilities plus a heatmap of the regions that drove the decision.

3

Auto-recommend

Detected problems flow into the biological advisor for a matched product.

Measured, not claimed

Benchmarked on the full held-out test set

10,878 leaf images the model never trained on. Every chart below is generated from the model's own predictions — 99.1% accuracy, 0.988 macro-F1.

38 by 38 confusion matrix, near-clean diagonal
Confusion matrix · 38 classesA near-clean diagonal. The only notable off-diagonal is the maize gray-leaf-spot vs northern-blight pair.
Per-class F1 for all 38 classes
Per-class F136 of 38 classes clear 0.95; most sit at 0.99+, so accuracy is not carried by a few big classes.
Accuracy per crop across 14 crops
Per-crop accuracyEvery one of the 14 crops scores at least 97.4%, and seven reach 100%.
Confidence distribution, correct versus incorrect
Confidence is a real trust signalCorrect predictions average 0.87 confidence, incorrect only 0.46, so low-confidence calls can be flagged.
Verified live

All seven problem statements, working end to end

Each one runs against the real backend, not a slide. The module that answers it is named alongside.

PS-01built

Application timing and readiness

A readiness band with the stage, method and dose to apply.

recommender.py
PS-02live

Climate-stress early warning

Composite score and warnings from satellite, weather and soil.

risk.py
PS-03built

Personalised advisor

Ranked biologicals with confidence and clear reasons.

recommender.py
PS-04Gemini

Multilingual chatbot

Grounded replies with session memory and an offline fallback.

chat.py
PS-05built

Field outcome logging

Season journal persisted to SQLite and retrievable per field.

learning.py
PS-06learns

Efficacy benchmark and learning

Real logged outcomes lifted Trichoderma efficacy 0.75 to 0.80.

learning.py
PS-07MSP

Yield attribution and ROI

Real MSP prices give ROI of 39.8% and net of Rs 37,013 per field.

economics.py

The loop, in one call

The /advisory endpoint runs risk, recommendation, advisory and economics together.

Try the full flow

Real data, no mock

The prototype runs on live public sources and falls back gracefully when a source is offline.

gee-sentinel2

Satellite vegetation

Sentinel-2 NDVI and NDMI series via Google Earth Engine.

live
open-meteo

Weather forecast

Temperature, humidity and rainfall driving the disease sub-score.

live
soilgrids-isric

Soil profile

SoilGrids 250 m pH and texture, resolved per location.

live
MSP 2024-25

Official prices

Minimum support prices with live Agmarknet for the ROI model.

official
gemini

Grounded advisory

Multilingual advice constrained strictly to the app's own facts.

grounded
outcomes.db

Field outcomes

Logged results retrain per-product efficacy every season.

closed loop
99.0%
EfficientNet-B4 validation accuracy
16
Biological products in the catalog
27/27
Backend tests passing
39.8%
Modelled ROI on a sample field

Run the full loop on a real field

Drop a pin, scan a leaf, read the risk, get a matched biological with its money case, and log the outcome. All in the browser.