AgriPrice AI — Helping farmers make smarter selling decisions with AI

The problem, our solution, the impact

Most Indian farmers decide when and where to sell with yesterday's mandi gossip. Prices for perishables swing 20–40 % within weeks; selling into a trough or hauling to the wrong market erases a season's margin.

❗

Problem

No forward view of prices — only today's rate, if that.
Market choice ignores transport, storage and spoilage.
Advice, when available, is in English and unexplained.
Distress sales right after harvest when supply peaks.

💡

Our solution

7 / 14 / 30-day price estimates with uncertainty, from a transparent ML pipeline.
SELL NOW / MONITOR / WAIT with reasons — shelf life and storage cost included.
Market comparison on net revenue, profit calculator, grounded advisor.
Five languages, voice in and out, and a marketplace to close the sale.

🌾

Economic impact

Better timing: avoid selling into predictable dips; hold storable crops when the expected gain beats holding cost.
Better routing: the best price is not always the best profit after transport.
Less waste: shelf-life-aware advice avoids holding perishables too long.
Direct buyer contact without commission.

No guaranteed-income claims: forecasts are estimates and the bundled data is synthetic.

Real model output — the five judge examples

Everything below was produced by the running app's own API (captured , demo dataset). Pick a case; change the quantity to see revenue scale. This is the same data the live dashboard shows.

Price history → estimated price₹ per kg · last 90 days + 30-day path

Actual modal priceEstimated price80 % prediction range

Sell today vs sell later

Quantity (kg): · revenue scales with quantity; holding cost includes crop-specific spoilage

Horizon table

Horizon	Estimate	80 % range	Change	Advice	Net revenue

Top model drivers (permutation importance)

Recommendation reasoning

Candidate models for this series

Model	MAE	RMSE	MAPE	R²	Skill	Dir. acc.

Compare markets —

Market	Today	7-day est.	Change	Confidence	Advice	Net today	Net in 7 d

Net = expected revenue − transport − storage − spoilage. Transport costs are per market and editable in the app; here they are 0, so ranking follows price × quantity after holding losses.

Price forecasts are estimates based on historical data and available market information. Actual prices may vary due to weather, supply, demand, transportation, government policies and other market conditions. The bundled dataset is synthetic and labelled as such on every screen; the ingestion layer accepts real Agmarknet / data.gov.in exports without code changes.

What the product does

A complete decision-support flow: crop → location → market → quantity → period → forecast → dashboard → compare → decide → sell.

📈

Price forecasting

7 / 14 / 30-day estimates with an 80 % prediction interval, trend direction and a High / Medium / Low confidence label — never presented as a guarantee.

🧭

Explainable SELL / MONITOR / WAIT

Rules over current price, estimate, trend, volatility, uncertainty, quantity, shelf life and storage cost. Every verdict comes with reasons and caveats.

⚖️

Compare markets on net revenue

Expected revenue − transport − storage − spoilage per market, so the best price is also the best profit.

🧮

Farmer profit calculator

Gross → spoilage → transport → storage → other costs → net; break-even price; waterfall chart.

🤖

AgriAdvisor

Rule-based assistant grounded in the same prediction numbers: "Should I sell my tomatoes today?", "What if price falls 10 %?", "Which market pays most?"

🎙️

Voice assistant

Speak in English, Tamil, Telugu, Kannada or Malayalam: the app runs predictions, compares markets, opens pages, switches language, pre-fills a listing — and reads the answer aloud.

🌐

Five languages

Whole UI, recommendation and explanation texts, advisor and voice. One type-checked dictionary per language.

🛒

Farmer marketplace

List produce at an AI-suggested asking price, buyers call / WhatsApp / send interest; owner view with inquiries; admin moderation. No payments handled.

🔍

Why this prediction?

Recent % change, seasonal pattern, arrivals, similar historical periods, model confidence, top model drivers.

🛠️

Admin & data layer

Upload CSV / Excel / JSON, column auto-mapping, cleaning report, data-quality view, retrain, model registry, prediction history. Agmarknet-compatible.

🔒

Security & honesty

Input validation, upload sanitising, token-protected admin, secrets via env vars, synthetic data labelled on every screen.

⚡

Demo mode

One-click examples (Tomato, Onion, Potato, Rice, Wheat) that render instantly — no API keys needed.

Voice assistant & five languages

Many farmers prefer to ask. The floating mic understands spoken commands in five languages and operates the app — then reads the grounded answer aloud.

How it works

Speech → text in the browser (Web Speech API, recognizer set to ta-IN / te-IN / kn-IN / ml-IN / en-IN). Only the transcript is sent to the server.
POST /api/voice/command parses intent + entities (crop, market, quantity, horizon, language) and returns a structured action: predict · compare · navigate · set_language · sell_online · chat · stop · repeat.
The UI executes it (runs the prediction, opens the comparison, switches language, pre-fills a listing) and speaks the reply.

Honest by design

Answers come from the same prediction services as the dashboard — nothing is made up.
No Tamil/Telugu/Kannada/Malayalam voice on the device? The text stays in that language, the English version is spoken, and the user is told why.
No microphone? A keyboard mode runs the same command pipeline.
No audio leaves the device; no cloud speech keys.

Try saying…

Tomato price in KoyambeduShould I sell 500 kg onion today?Which market is best for potato?கோயம்பேடு தக்காளி விலை என்னಕೋಲಾರದಲ್ಲಿ ಟೊಮೇಟೊ ಬೆಲೆ ಎಷ್ಟುSell my 500 kg tomatoes onlineSwitch to TeluguOpen the calculator

Languages

🇬🇧 Englishதமிழ்తెలుగుಕನ್ನಡമലയാളം

Whole UI, recommendation and “why” texts, advisor and voice. One type-checked dictionary file per language — the compiler rejects a missing key. Numbers and dates use the matching Indian locale.

**Voice assistant**“Should I sell 500 kg onion today?” → prediction runs, answer is read aloud.

Tamil voice — **Tamil**Spoken Tamil → Tamil dashboard + Tamil speech.

Farmer marketplace — from insight to sale

The forecast tells the farmer when to sell; the marketplace helps with to whom. Currently 8 active sample listings · 17,850 kg.

🏷️

AI-suggested asking price

Pre-filled from today's mandi modal price and the 7-day estimate, with the range, trend and confidence shown — a starting point, not advice.

📞

Direct buyer contact

Call, WhatsApp, or an “I'm interested” form with offer and quantity. Phone numbers are masked unless the seller opts in.

🗂️

Owner tools

One-time edit key, inquiries list, mark as sold, remove, WhatsApp share. Admin moderation tab with full details.

🛡️

Safety by design

No payments, escrow or logistics; on-page safety notice; rate limits, input sanitising, Indian mobile validation.

ML methodology & evaluation

A real, reproducible pipeline — no random numbers. Every number in the UI traces back to a trained, evaluated model stored with its metrics.

1 · Clean

Drop rows missing essentials, de-duplicate, fix swapped min/max
Flag abnormal prices (rolling MAD outliers), chronological sort, ₹/quintal → ₹/kg
Daily aggregation with a transparent cleaning report (28,921 → 28,086 rows)

2 · Features

Calendar: day, week, month, year, day-of-week, seasonal sin/cos
Lags 1–30, rolling mean/min/max/std, % changes, moving averages
Arrivals, volatility, horizon feature h ∈ 1…30 (direct multi-horizon), log-return target

3 · Models

Baseline: naive last value · Linear (Ridge)
Improved: Random Forest · HistGradientBoosting · XGBoost
Per crop × market: hold-out last 120 days, select by RMSE, fall back to naive unless beaten by > 2 % — honesty over show
80 % intervals from hold-out residual quantiles; LSTM/Prophet deliberately not used for show

Average over all 25 series (hold-out, horizons 1–30)

Model	MAE ₹/kg	RMSE	MAPE	R²	Skill vs naive	Best in
Naive (last price)	2.58	3.38	8.9 %	0.19	—	1 / 25
Linear Regression (Ridge)	1.78	2.27	6.1 %	0.43	+33 %	7 / 25
Random Forest	1.75	2.36	6.0 %	0.40	+30 %	14 / 25
Gradient Boosting (HistGB)	1.88	2.54	6.5 %	—	—	2 / 25
XGBoost	1.84	2.47	6.3 %	—	—	1 / 25

Selected models: mean MAPE 6.0 %, mean skill +42 %; 24 of 25 series beat naive by more than 2 % RMSE. Full per-series table in the README and notebook.

By crop (selected model)

Crop	MAPE	Skill	Model
🧅 Onion	4.9 %	+79 %	Random Forest (all 5 markets)
🥔 Potato	5.7 %	+21 %	Ridge (all 5 markets)
🌾 Rice (Paddy)	1.6 %	+17 %	Ridge / RF; Thanjavur honestly falls back to naive
🍅 Tomato	12.7 %	+63 %	HistGB, Ridge, XGBoost (per market)
🌾 Wheat	2.5 %	+27 %	Random Forest (all 5 markets)

Tomato is the hardest (volatile, perishable) and is where skill over naive is largest; Rice/Thanjavur is the one series where the honest naive fallback wins.

Decision engine (explainable rules)

Spoilage = daily loss × days (tomato 1 %/day, shelf life 7 d · onion 0.25 %/day, 60 d · potato 0.2 %/day, 75 d · rice/wheat 0.02 %/day). Net expected change = (estimate × qty × (1 − spoilage) − storage − revenue now) ÷ revenue now. WAIT when net gain ≥ 5 %, downside of the 80 % band better than −8 % and confidence not low · SELL NOW when net ≤ −2 % or past shelf life with no gain · otherwise MONITOR. Every verdict lists its reasons and caveats and is never framed as certainty.

Architecture & tech stack

React + TypeScript + Tailwind + Recharts in front; FastAPI + Pandas/NumPy/scikit-learn/XGBoost behind; SQLite (PostgreSQL-ready via DATABASE_URL); joblib model registry. Single process in production: FastAPI serves the built frontend.

frontend/

pages: Landing · Predict · Compare · Calculator · Advisor · Marketplace · Admin
components: charts, results, forms, AgriAdvisor, voice/
voice/useSpeech.ts — Web Speech wrapper
i18n: en · ta · te · kn · ml (one Dict type)

backend/app/

api/routes: public · voice · marketplace · admin
ml: preprocessing · features · models · forecaster · recommender · explainer
services: prediction (registry + cache) · advisor · voice · marketplace
core: config (env) · security (admin token, upload validation)

data & models

data/generate_sample_dataset.py → sample/agri_prices_demo.csv (synthetic, seeded)
ingestion: CSV · Excel · JSON · data.gov.in API, auto column mapping
models/<crop>__<market>.joblib + metrics in SQLite
notebooks/01_eda_and_model_comparison.ipynb

quality

33 pytest tests (API, ML, voice, marketplace)
Playwright E2E: voice, 5 languages, marketplace (26 steps)
Docker compose, Makefile, typed API client, tsc strict

Browser ──/api/predict──▶ FastAPI ──▶ prediction_service.get_or_train() ├─ series_frame() from SQLite ├─ model registry + joblib cache (fingerprint-validated) ├─ forecaster.forecast_series() → 30-day path + 80 % intervals ├─ recommender.recommend() → SELL / MONITOR / WAIT + reasons └─ explainer.explain() → "why" bullets + drivers Mic ──(browser STT)──▶ /api/voice/command ──▶ intent + entities ──▶ same services ──▶ action + reply ──▶ UI + TTS

REST API

Method	Route	Purpose
`GET`	`/api/crops · /api/markets · /api/locations`	catalogue with history coverage
`POST`	`/api/predict`	forecast + recommendation + explanation + model metrics
`POST`	`/api/compare-markets`	net revenue by market after transport & holding costs
`POST`	`/api/calculate-revenue`	profit calculator arithmetic
`GET`	`/api/price-history · /api/model-metrics`	cleaned history, registry metrics
`POST`	`/api/chat`	AgriAdvisor (grounded, 5 languages, TTS-ready text)
`POST`	`/api/voice/command`	transcript → structured UI action + spoken reply
`GET/POST`	`/api/marketplace/…`	price suggestion, listings, inquiries, status, stats
`POST`	`/api/upload-data 🔒 · /api/admin/… 🔒`	ingestion, dataset, retrain, moderation (X-Admin-Token)

Explicit errors: 404 unknown crop/market · 422 “Not enough historical data is available for this crop and market. Please select another market or prediction period.” · 401 admin · 400/413 bad or oversized upload.

Screenshots

Captured from the running app with Playwright (1366 px and 390 px). Click to enlarge.

**Landing page**Hero, impact cards and the economic-impact section.

**Prediction dashboard**Tomato · Koyambedu · 500 kg · 7 days - five summary cards and the explainable recommendation.

**A WAIT case**Onion · Tiruchirappalli · 2000 kg · 14 days: +16 % expected with high confidence.

**Full prediction page**Charts, 7/14/30-day table, 'Why this prediction?' and the model comparison.

**Compare markets**Ranked by net revenue after transport and holding costs - not by headline price.

**Farmer profit calculator**Sell today vs later, waterfall to net profit, break-even price.

**AgriAdvisor**Answers grounded in the same prediction numbers (what-if, best market, revenue).

**Voice in Tamil**Spoken Tamil command → Tamil dashboard and Tamil speech (recognizer ta-IN).

**Telugu**The complete dashboard in తెలుగు.

**Kannada**The complete dashboard in ಕನ್ನಡ.

**Malayalam**The complete dashboard in മലയാളം.

**Honest voice fallback**No Malayalam voice installed → text stays Malayalam, answer read in English, user told why.

**Marketplace - sell**Form pre-filled by voice, AI asking-price suggestion from today's mandi price and the 7-day estimate.

**Marketplace - browse**'vs mandi' badges, call / WhatsApp / I'm interested.

**Marketplace - my listings**Interested buyers with offers; mark as sold.

**Admin**Dataset browser, cleaning report, upload, models, data quality, prediction log.

**Phone layout**Voice panel in Kannada over the prediction page (390 px).

The 90-second demo

Open the app → View Demo (Demo Mode banner states the data is synthetic).

Click 🍅 Tomato · Koyambedu · 500 kg · 7 d — the form fills and the prediction runs.

Read the five cards: current price, estimated price, expected change, recommendation, additional revenue.

Scroll: forecast chart with the 80 % band, price history, sell-today-vs-later bars, the 7/14/30-day table.

“Why this prediction?” and the model table (naive vs linear vs trees with skill %).

Click 🧅 Onion · Tiruchirappalli · 2000 kg · 14 d → a 🔵 WAIT case (+16 %, high confidence, ₹9.7 k more net).

Compare Markets → add a transport cost → the best market is ranked by net revenue, not headline price.

Ask AgriAdvisor “What happens if the price falls by 10 %?” — the answer uses the same numbers.

Switch to தமிழ் — everything, including the recommendation, is translated.

Tap Talk to AgriPrice: “Which market is best for potato?” → comparison opens and is read aloud; “Sell my 500 kg tomatoes online” → marketplace pre-filled with an AI asking price.

Run it locally

git clone <repo> && cd agriprice-ai make setup # pip + npm install make demo # data → seed → sample listings → train 25 models → build UI make backend # http://localhost:8000 (API docs at /api/docs)

Or ADMIN_TOKEN=secret docker compose up --build. Voice input needs https:// or localhost (browser microphone policy).

Limitations & next steps

Synthetic demo data — accuracy must be re-validated on live Agmarknet feeds (ingestion is ready).
No weather / policy / export-ban features yet; horizons beyond ~2 weeks for volatile vegetables carry wide intervals.
Advisor and voice are rule-based: dependable within their intents, not open-ended chat.
Marketplace is a moderated classifieds board — no identity verification, payments or logistics.
Next: live data sync, weather & festival features, quantile boosting, SMS/IVR, OTP-verified sellers, PostgreSQL + scheduled retraining.

The problem, our solution, the impact

Problem

Our solution

Economic impact

Real model output — the five judge examples

Sell today vs sell later

Horizon table

Top model drivers (permutation importance)

Recommendation reasoning

Candidate models for this series

Compare markets —

What the product does

Price forecasting

Explainable SELL / MONITOR / WAIT

Compare markets on net revenue

Farmer profit calculator

AgriAdvisor

Voice assistant

Five languages

Farmer marketplace

Why this prediction?

Admin & data layer

Security & honesty

Demo mode

Voice assistant & five languages

How it works

Honest by design

Try saying…

Languages

Farmer marketplace — from insight to sale

AI-suggested asking price

Direct buyer contact

Owner tools

Safety by design

ML methodology & evaluation

1 · Clean

2 · Features

3 · Models

Average over all 25 series (hold-out, horizons 1–30)

By crop (selected model)

Decision engine (explainable rules)

Architecture & tech stack

frontend/

backend/app/

data & models

quality

REST API

Screenshots

The 90-second demo

Run it locally

Limitations & next steps

Team