Planotell

Problem

Working out what a plant is — and then what is wrong with it — usually means guesswork or scattered forum threads. And even with an answer, plants die of the follow-through: the missed watering, the treatment abandoned halfway. People want one place that tells them what the plant is, whether it is healthy, what to do about it — and then reminds them to actually do it.

Solution

Planotell (launched as PlantPal) turns a single photo into a full care loop. The species is identified and its health checked in the same scan; a reasoning model turns the raw classification into a personalised watering, feeding and treatment plan, and push reminders keep it on schedule. Every diagnosis comes with a step-by-step treatment plan tracked through to recovery, and a chat assistant that knows the plants you actually own answers follow-up questions, streamed as it thinks. Identification runs asynchronously over Kafka so a slow model never blocks the app, and a provider-agnostic AI layer routes each task across several vision and reasoning models — hosted and local — with the user able to choose. Every identification also enriches a shared botanical knowledge base, so advice improves for everyone. The app speaks English, French and Arabic (right-to-left included), and it is live (beta) at planotell.com — frontend on Vercel, backend on its own hardened VPS. A scan taken offline waits on the device and uploads when you are back online, and uploaded photos are stripped of GPS data before storage or AI use.

Features

• Species identification and health check from one photo
• Personalised care plans with web-push reminders
• Disease detection with step-by-step, trackable treatment plans
• Plant-care chat that knows your garden, streamed in real time
• Multi-model AI — Claude, GPT-4o, PlantNet, DeepSeek and a local fallback
• Shared species knowledge base that grows with every scan
• English, French and Arabic, with right-to-left layout and read-aloud
• Installable PWA plus an administration console
• Offline scan queue — scans taken without a connection upload automatically later, with retry and discard
• First-run guidance: a skippable intro, photo tips and an upload fallback
• Plant groups by room, bed or greenhouse, with one-tap group watering or feeding
• Plant journal for repotting, pruning, flowering, fruiting and propagation
• "Was this right?" feedback on identifications, reviewable by an admin
• Photo timelines per plant and per species, ordered by the date taken
• Privacy by design — GPS coordinates stripped from every uploaded photo; only a coarse region and light/placement are kept, never a precise location
• Self-service data export — plants, care log, scans and journal as JSON and CSV, plus photos
• Crawlable English and French public pages and transactional email (verification, recovery)
• Task-driven delivery with a SonarQube quality gate on every pull request

Architecture

1. Angular PWA
2. Spring Boot API
3. Kafka identification pipeline
4. Vision + reasoning models
5. PostgreSQL · Redis
6. Care plans, reminders & web push

Tech

• Angular
• Spring Boot
• Java 21
• PostgreSQL
• Redis
• Kafka
• Vision Models
• LLMs
• Server-Sent Events
• Web Push
• Docker
• Vercel
• SonarQube
• Playwright
• Testcontainers

Challenges

• Combining two very different AI models — a vision classifier and a reasoning LLM — into one coherent answer.
• Keeping the conversation anchored to the actual image and classification rather than letting the LLM drift.
• Living with real AI providers in production: rate limits, empty or truncated replies and slow models — handled with an asynchronous pipeline, honest retry timing and fallbacks instead of a spinner that never ends.
• Three languages over AI-generated content: each section keeps its own saved translation and is only translated on demand, so switching language never silently re-bills the whole garden.

Lessons

• Prompt engineering matters as much as model choice: the product only feels intelligent when the reasoning step is tied to real classifier output and conversation memory.
• A provider-agnostic AI layer pays for itself — the models underneath changed several times; the product did not.
• Shipping is its own discipline: the first production boot needed six layered fixes, and a quality gate (SonarQube, blocking auth test suites) is what keeps the next release from needing them again.

Future work

• Broaden the species database
• On-device / offline recognition
• Graduate the experimental "atlas" interface — a second, exploratory frontend over the same API — into the main app