Inspiration
Plant care advice today is mostly static, generic, and reactive. Most apps analyze a single image and return one-time suggestions without learning whether those suggestions actually worked. We wanted to explore how Gemini 3 Pro could be used not as a one-shot analyzer, but as an orchestrator that continuously reasons about plant health over time, learns from outcomes, and adapts its decisions autonomously.
What it does
PlantOps AI is an orchestrated, multimodal system that continuously monitors plant health using image-based reasoning over time. Users periodically upload plant images, and the system tracks visual changes, compares them against historical context, and reasons about cause-and-effect relationships such as watering patterns, light exposure, and visible stress indicators.
Instead of providing static advice, PlantOps AI maintains a persistent plant profile, evaluates whether past recommendations improved or worsened plant health, and autonomously adjusts future care strategies.
How we built it
We built PlantOps AI around Gemini 3 Pro acting as the central orchestrator, not just a responder. The system follows a multi-step reasoning loop:
- Multimodal analysis of uploaded plant images
- Comparison against historical plant states stored in long-term memory
- Hypothesis generation about plant stress or improvement
- Tool-driven evaluation and reasoning refinement
- Adaptive recommendation planning for future observation cycles
Gemini’s large context window enables full lifecycle reasoning across multiple uploads, while its tool-calling and thinking levels allow the system to plan, execute, verify, and adjust decisions autonomously.
Challenges we ran into
One major challenge was avoiding baseline image analysis and prompt-only flows. Designing the system as an adaptive, stateful agent required careful orchestration of memory, reasoning steps, and verification logic. Another challenge was ensuring the system focuses on observable plant care signals without crossing into medical or diagnostic claims.
Accomplishments that we're proud of
- Built a true orchestrated agent rather than a single-prompt application
- Implemented time-based reasoning and outcome evaluation
- Designed a system where Gemini’s reasoning is essential and non-replaceable
- Demonstrated autonomous adaptation instead of static recommendations
What we learned
We learned that orchestration is less about complex prompts and more about designing feedback loops where the model reasons over history, verifies outcomes, and adapts its behavior. Gemini 3 Pro excels when treated as a long-running decision-maker rather than a one-time generator.
What's next for PlantOps
Next, we plan to expand PlantOps into a long-running Marathon Agent that can autonomously schedule observation reminders, integrate environmental signals like weather and sunlight trends, and reason over weeks or months of plant health evolution. We also aim to introduce predictive alerts and proactive care planning based on learned plant behavior patterns.
Log in or sign up for Devpost to join the conversation.