EcoBloom | Devpost

Inspiration 🌸

We started with the Deloitte Green AI challenge, how can AI be resource-efficient while solving environmental problems? That led us to hardware: what if we built a real-time embedded system that makes energy waste visible and emotional rather than just showing numbers on a dashboard? Enter EcoBloom: a plant that literally reacts to your household's energy habits in real time.

The core insight: people waste energy without realizing it. A dashboard telling you "lights on 23 minutes" gets ignored. A plant visibly wilting because you left the lights on, you can't ignore that.

What it does

EcoBloom uses computer vision to detect room conditions (dark, lit with people, lit without people) and translates those states into a "mood score" that drives real-time environmental feedback:

The Plant (Hardware): Green and yellow LEDs fade in real time, as mood drops from 100 (healthy, full green) to 0 (dying, full yellow), the colors crossfade smoothly. It's a physical, emotional representation of energy waste.

The Chatbot (AI): Ask "why are you dying?" and EcoBloom, powered by Backboard, answers with reasons grounded in your actual sensor log. Not generic advice, specific to what happened in your home ("lights were on for 145 seconds with no one around").

The Dashboard (Web): A responsive website shows live mood/HP, event history, plant graphics that match the physical LED state, and the conversational chatbot.

The Backend (Real-time Logic): Built on QNX Raspberry Pi, the system runs camera detection, logs state transitions with precise timing, and exposes a clean API for all other components.

How we built it

Tech Stack:

Backend: Python Flask (mood scoring, event logging, state machine) + QNX Raspberry Pi (real-time GPIO control)
AI/Reasoning: Backboard.io (LLM orchestration, grounding answers in event logs)
Frontend: React + HTML/CSS (responsive dashboard)

Key design decisions:

Stateful mood scoring instead of threshold-based: tracks duration of bad states so brief light-ons don't trigger false alarms, but sustained waste gets caught immediately
Local Backboard inference means answers ground in real data, not generic LLM hallucination
QNX + real-time GPIO ensures the LEDs respond instantly (sub-100ms) rather than polling delays, critical for a believable, reactive demo
Mock camera module lets the team test the full pipeline before camera detection code exists

Challenges we ran into

Hardware wiring: Getting male-to-female jumper wires and correctly wiring LEDs to GPIO pins took longer than expected. We're still finalizing the physical LED rig, but the code is complete and tested.

Real-time coordination: Coordinating three separate systems (camera detection, mood server, LED controller, web frontend) meant careful API contract design upfront, we used mocked components early so no one was blocked.

Backboard API learning curve: First time using Backboard; had to learn model routing and thread management, but the platform's straightforward REST API made it fast to integrate once we understood the model.

Sensor noise: Camera detection can flicker on dark rooms or under bad lighting. We added temporal smoothing (duration thresholds) to avoid false state transitions.

Accomplishments that we're proud of

✅ Full end-to-end pipeline working: from camera → mood logic → event log → Backboard reasoning → website display, all integrated and tested

✅ Backboard grounding: the chatbot doesn't just generate text, it reads your actual event log and grounds answers in real data ("the lights were on for 145 seconds...", that's a real number from your home)

✅ Smooth LED crossfade: the physical feedback isn't jarring on/off — it's a beautiful, continuous gradient from green to yellow as mood decays

✅ QNX real-time system: chose QNX specifically for the Deloitte track and because real-time guarantees matter for a reactive device; the system runs deterministically

✅ Demo-ready in 36 hours: even with hardware incomplete, the full software pipeline is live, tested, and repeatable

What we learned

Real-time systems matter: Building on QNX forced us to think about latency, GPIO timing, and thread priorities, totally different from typical web development, and it's why the LEDs feel responsive.

Hardware is significantly harder than code: Wiring, soldering, debugging physical connections takes way longer than we thought. Respect to hardware engineers.

API orchestration: Using Backboard to route to LLMs, grounding responses in data, and coordinating multiple microservices taught us how to build scalable AI systems beyond just calling ChatGPT.

Behavioral design beats data design: A plant that visibly dies is 1000x more impactful than a dashboard chart showing "energy waste percentages." Emotion drives behavior change.

What's next for EcoBloom

Finish LED hardware wiring and integrate the live physical demo (flower)
Deploy the Flask backend to a cloud platform (Render/Railway) so it's accessible over network
Expanding number of sensors