title	Plotwise
emoji	🌾
colorFrom	green
colorTo	gray
sdk	docker
app_port	8080
pinned	false

Plotwise

Agricultural intelligence platform for Nagaland, India: on-device crop-disease detection, market prices with a live mandi feed, planting calendars, weather, and a government-scheme finder — built on the state's real 2023-24 production records. FastAPI + TensorFlow backend, framework-free HTML/JS frontends, Capacitor Android app.

Live demo: https://limajmr-plotwise.hf.space · Author: Limawapang L Jamir

I audited my own test set and found 427 images I had already trained on — how a contaminated split inflated this project's headline accuracy, why the corrected number matters less than the 43.6% it led me to measure, and what the confidence system does not protect against.

---

What's technically unusual here

1. Inference ships to the phone. The trained CNN (EfficientNetB0, 24 classes) is converted to a 9 MB TensorFlow.js graph model (frontend/src/tfjs-model/, conversion metadata recorded in model.json: TensorFlow.js Converter 4.22.0 from TF 2.17.0) and bundled inside the Android APK — capacitor.config.ts has no remote server.url; scripts/build_bundle.sh packs the UI, model, and reference data into the app. In the APK, leaf diagnosis therefore runs with zero connectivity, behind a bilingual English/Nagamese UI (the diagnosis and treatment text itself is currently English-only); offline-ai.js is a JS port of the server's crop-aware decision layer, and offline detections queue locally and sync to the server's surveillance heatmap when a connection returns.

In the browser, that same code path is deliberately switched off: running TensorFlow.js requires 'unsafe-eval', which this app's Content-Security-Policy (main.py CONTENT_SECURITY_POLICY) does not grant, so the web PWA sends photos to the server endpoint instead. The APK is unaffected because it serves no CSP header. The trade — a hardened header on the public web, full on-device inference in the app — is deliberate, and offline-ai.js ready() detects the disabled runtime and falls back rather than failing. Measured, with the reasoning, in CHANGELOG 2.5.2.

2. The accuracy claims are measured — including the bad ones. Models trained on lab datasets are known to collapse on real field photos. Instead of quoting the lab number, eval/ measures the deployed inference path on in-the-wild images and reports the collapse, what the decision layer recovers, and where the confidence system fails. See the table below.

3. An evaluated agent layer, not a chatbot wrapper. The assistant can run as a tool-use loop over the backend's data functions, with local-model routing (local 7B via Ollama, no API cost) and escalation to claude-opus-5 — the routing is built, though the measured numbers so far cover the local backend only. Graded by a committed 40-question eval with ground truth derived from the dataset. See below.

---

Generalization: lab accuracy vs. field accuracy

The same deployed inference code (imported from backend/src/main.py, not reimplemented) evaluated on held-out lab images (PlantVillage) and on real field photographs (PlantDoc):

	PlantVillage held-out, clean (n=7,654)³	PlantDoc field images (n=1,642)
A — raw model argmax, top-1	87.7%	27.3%
B — + crop-aware pass, top-1	91.9%¹	43.6%
C — + abstention: coverage	89.9%	48.4%
C — selective accuracy (when it answers)	96.4%	45.7%
OOD images routed to "uncertain"²	—	75.7% same-crop · 90.6% foreign-species

¹ n=7,588 (66 images with no crop label excluded from the crop-aware arm), so the lab delta below compares across the two denominators. ² Computed on the 936 unmapped OOD images (305 same-crop, 631 foreign-species), not the 1,642 in the column header. ³ A 2026-08 content-hash audit found 427 of the 8,081 test images byte-identical to training files (residue of an earlier dataset prep); they are excluded from every number here, the runner re-excludes them on every run, and the leaked subset's 98.7% score is reported in eval/RESULTS.md §1a as the memorization signature it is.

python eval/plantvillage/run_baseline.py   # left column
python eval/plantdoc/download.py           # fetch PlantDoc (CC BY 4.0)
python eval/plantdoc/run_eval.py           # right column

Honest reading: the crop-aware pass genuinely helps everywhere (+4.2 pts lab, +16.3 pts field), but field accuracy is 43.6%, not 91.9% — and under domain shift the confidence system stops working as a safety net (on answered field images the error rate is 54.3%; in-domain it is 3.6%). The part that does survive is OOD routing: photos of untrained diseases and foreign species are mostly sent to "uncertain" instead of answered confidently. Full analysis, per-class breakdowns, risk–coverage curves, and dataset caveats: eval/RESULTS.md.

---

Architecture

flowchart TB
    subgraph train["Model pipeline (offline, GPU)"]
        PV["PlantVillage dataset"] --> TRAIN["ml/train_disease_model.py<br/>EfficientNetB0 · 24 classes"]
        TRAIN --> H5["disease_model.h5"]
        H5 -->|"tensorflowjs_converter 4.22.0"| TFJS["frontend/src/tfjs-model/<br/>9 MB graph model"]
    end

    subgraph server["FastAPI backend (backend/src/main.py)"]
        CSV["576-record CSV<br/>Director of Agriculture, Nagaland 2023-24"] --> API["REST API · PDF/CSV export · chatbot"]
        MANDI["data.gov.in Agmarknet feed<br/>12 h cache · 7-day stale guard"] --> API
        METEO["Open-Meteo weather"] --> API
        API --- AGENT["backend/src/agent.py<br/>tool-use loop (PLOTWISE_AGENT=1)"]
        AGENT -->|"simple lookups"| OLLAMA["qwen2.5:7b via Ollama<br/>(local, $0)"]
        AGENT -->|"multi-step escalation"| CLAUDE["claude-opus-5"]
    end

    subgraph clients["Clients"]
        WEB["Desktop SPA<br/>index.html"]
        PWA["Mobile PWA<br/>mobile.html + sw.js app-shell cache"]
        APK["Android APK (Capacitor)<br/>UI + model + data bundled"]
    end

    H5 --> API
    API <--> WEB
    API <--> PWA
    TFJS -->|"scripts/build_bundle.sh"| APK
    APK -.->|"offline queue → sync on reconnect"| API

Loading

---

Data provenance

• Crop records — 576 records: "District-wise Area, Production and Yield (Achievement) 2023-24", Director of Agriculture, Nagaland, Kohima (verified 09-May-2024, checked 30-Jul-2024). 16 districts, 44 crops. This is the ground truth for the dashboard, the yield API, and most of the agent-eval questions.
• Market prices — live daily mandi rates for Nagaland APMCs from the official data.gov.in Agmarknet feed (12-hour cache, 7-day stale guard) where mandis report; MSP 2023-24 (CACP) + Agmarknet anchors as clearly-labelled indicative fallback. A live rate is never invented — no key, no feed, or stale data means the row says "Indicative".
• Planting calendar — sow/harvest windows checked against the official Nagaland Crop Calendar (Department of Agriculture, Government of Nagaland, web-published Jan 2025); zone lists derived from the 2023-24 CSV by scripts/audit_calendar_vs_csv.py and drift-guarded by the test suite.
• Weather — Open-Meteo (no API key), GPS coordinates for all 16 district HQs: Kohima, Tseminyu, Phek, Mokokchung, Tuensang, Noklak, Shamator, Mon, Dimapur, Niuland, Chumoukedima, Wokha, Zunheboto, Peren, Kiphire, Longleng.
• Disease model — PlantVillage dataset (Penn State). Field eval: PlantDoc — Singh et al., CoDS-COMAD 2020, DOI 10.1145/3371158.3371196, CC BY 4.0.

---

Agent layer

backend/src/agent.py replaces keyword intent-detection with a model-driven tool-use loop over six backend functions (get_yield, get_price, get_weather, compare_districts, get_planting_calendar, find_schemes). Feature-flagged (PLOTWISE_AGENT=1); the legacy intent path stays as the default and as the fallback when no backend is reachable. A rule-based router sends simple lookups to a local model and escalates comparisons and multi-step questions to claude-opus-5; every turn is logged as JSONL (tools, arguments, latency, token cost). Numeric answers must trace to tool results — no data means a refusal, not a guess.

Graded by a 40-question eval: 25 answerable, with ground truth derived from committed sources — the 576-record CSV plus the backend's price, calendar, and scheme tables (one live-weather question is graded on tool selection only, hence 24 graded below); 10 unanswerable, where refusal is correct; and 5 ambiguous, where a clarifying question is correct:

Metric	local backend (qwen2.5:7b)
Tool-selection accuracy	100% (25/25)
Answer correctness (24 graded)	70.8%
Refusal accuracy (10 unanswerable)	90%
Clarification rate (5 ambiguous)	20%
Mean latency	1.7 s
API cost	$0

python eval/agent/run_eval.py --backend local

The local 7B model picks the right tools and refuses well, but tends to answer one interpretation of an ambiguous question instead of asking, and fumbles multi-step arithmetic — exactly the traffic the router escalates. Claude and routed columns are pending an ANTHROPIC_API_KEY. Grader design, iteration history (recorded honestly, including the bugs that inflated early numbers), and per-question outputs: eval/README.md.

---

What it does

• Disease detection — photograph a leaf → prediction with three-tier confidence (confident ≥70% / low 55–70% / uncertain — the app says "not sure" instead of guessing) and treatment guidance; works offline on the phone. Crops outside the model fall back to a curated knowledge base.
• Market prices — live vs. indicative labelling on every surface (UI, CSV export, PDF report, chatbot).
• Planting calendar — statewide sow/harvest windows checked against the official Nagaland Crop Calendar, filtered per district to the crops its 2023-24 records support, with live seasonal status.
• Scheme finder — crop → matching government schemes with eligibility notes.
• Weather — current + 7-day forecast with farming advisories, all 16 district HQs.
• Analytics dashboard — district-wise area/production/yield, disease surveillance heatmap with audit trail, one-click district PDF report.
• Assistant — English replies behind a bilingual EN/Nagamese chat UI; agent mode behind a flag.

---

Quick start (local)

Requires Python 3.11 — TensorFlow 2.17 has no wheels for newer Pythons.

git clone https://github.com/Limajmr8/plotwise.git
cd plotwise
python3.11 -m venv venv
source venv/bin/activate          # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn backend.src.main:app --port 8000

Open http://127.0.0.1:8000 (desktop UI) or /mobile (mobile UI).

⚠️ Do not use --reload on Windows — it crashes the TensorFlow subprocess. Keep the pinned versions in requirements.txt (TF 2.17.0 / Keras 3.10.0) — newer versions cannot load the model.

Optional configuration via .env (see .env.example): database path, CORS origins, live-price API key, agent flag and backends.

Docker

docker build -t plotwise .
docker run -p 8080:8080 -e PLOTWISE_SEED_ON_EMPTY=1 plotwise

---

API endpoints (selected)

Method	Endpoint	Description
GET	/health	Health check
GET	/api/status	App status, version, data summary
GET	/districts · /crops	Reference data (16 districts, 44 crops)
GET	/prices	Market prices (filters: crop, district)
GET	/calendar	Planting calendar with live status (filters: district, crop)
POST	/schemes	Scheme matching by crop
POST	/disease/detect	Leaf image upload → ML prediction (multipart)
GET	/api/weather	Live weather + 7-day forecast per district
POST	/api/chat	Assistant — agent loop when PLOTWISE_AGENT=1, keyword intents otherwise
GET	/dashboard/yield · /dashboard/disease-heatmap	Analytics
GET	/api/export/report	District PDF report
GET	/api/export/prices · /api/export/yield	CSV exports

Interactive docs at /docs when running locally.

curl -X POST http://localhost:8000/disease/detect \
  -F "file=@leaf.jpg" \
  -F "crop=Potato" -F "district=Kohima"

---

Android app

Capacitor bundles the mobile UI, the TF.js model, and the reference data into the APK (npx cap sync android after scripts/build_bundle.sh), so detection and the data tools work with no network; live features call the hosted backend. GitHub Actions builds a debug APK on every push to main.

---

Tests

pytest tests/ -q     # also runs in CI on every push

Covers every endpoint, the ML confidence tiers, the live-mandi cache (hit/miss/stale/failure-backoff), the weather proxy (mocked Open-Meteo), the agent loop (scripted fake backend — no network), a security-regression suite (XSS sanitisation, CSV-injection guards, filename injection, input validation), exports, edge cases, and data-integrity guards (the offline bundle must match the backend constants; calendar zones must match the CSV). Tests run against an isolated temporary database with the live feed disabled.

The backend is security-hardened from a whitebox audit with live exploitation: server-side input sanitisation + output escaping, a Content-Security-Policy, and injection guards — all under regression tests — plus per-client rate limiting and CVE-patched dependency pins.

---

Known limitations

• Field accuracy is 43.6%, and abstention is not a reliable safety net under domain shift — on field images the system still answers 48% of the time with a 54.3% error rate on those answers. The eval exists precisely to keep this number visible. Mitigations that would actually move it (fine-tuning on field imagery, adding the missing healthy/mosaic classes, confidence calibration on field data, background-robustness augmentation or leaf segmentation) are discussed in eval/RESULTS.md.
• Same-crop OOD is the dangerous case: a photo of an untrained disease on a covered crop is routed to "uncertain" only 75.7% of the time (worst: tomato mosaic virus, 13% routed — it usually gets a confident wrong answer).
• Class coverage: 24 classes over 9 crops. Nagaland staples such as paddy and ginger have no model coverage and fall back to a knowledge base.
• Mandi coverage is sparse: Nagaland APMCs report irregularly and mostly horticulture/perishables, so many price rows remain MSP-anchored "Indicative" even with a live key configured.
• Single data year: all yield data is 2023-24; no trends across years.
• Planting windows are statewide simplifications: the official Nagaland Crop Calendar documents different windows per district and farming system (jhum / terrace / plains); the app carries one modal [start, end] window per crop. Zone lists are derived from the 2023-24 CSV and drift-guarded by tests, but documented per-district exceptions (plains potato sown Oct–Nov, Tuensang's autumn potato, Mon's Aug–Sep rabi-maize cycle) are not modelled — see CHANGELOG 2.3.1 for the full audit trail.
• PlantDoc caveats: label noise, a known apple train/test inconsistency, and some very small classes — documented per-decision in eval/plantdoc/class_map.yaml.
• Agent eval grading is heuristic (deterministic keyword/numeric matching, no LLM judge) — cheap and reproducible, but crude; audit any grade from the committed per-question JSONL.

---

Author

Limawapang Jamir | B.Tech CSE, Bennett University (2020–2024) From Mokokchung, Nagaland 📧 limawapang8@gmail.com | LinkedIn | GitHub