BioSphere

Inspiration

Bioreactors are everywhere, from brewing and pharmaceuticals to algae farms and lab-grown food, but most of them are either hand-monitored by a human with a clipboard or locked behind six-figure industrial SCADA systems. We wanted to see if a small team could bridge that gap: a low-cost, camera and sensor-driven reactor that anyone can run from a browser, with an AI copilot watching over it.

What it does

JERM is a smart bioreactor platform that combines live hardware telemetry, computer-vision turbidity analysis, and an AI assistant into a single real-time dashboard.

Live sensor monitoring — temperature, humidity, pH, and light stream from an Arduino over serial at 2-second intervals.
Camera-based turbidity — the browser's webcam measures sharpness, color, and brightness to estimate how cloudy the culture is, with a one-click calibration baseline.
Health scoring — a blended rules + ML model fuses sensor readings and image features into a single health score with trend and anomaly risk.
AI assistant (JERM AI) — a Cerebras-powered chat sidebar that has access to the live reactor state, answers questions and suggests parameter adjustments.
Remote actuation — toggle the stirrer motor from the dashboard or via a physical button on the device; both stay in sync.
Auto-recommendations — the backend generates plain-English nudges ("raise temperature by 0.5 °C", "pH trending low") based on the current component scores.

How we built it

Hardware: Arduino with a DS18B20 temperature probe, pH sensor, light/humidity sensors, servo-driven stirrer, and a tactile override button. Firmware speaks a simple line-based protocol over serial.
Backend: Python + FastAPI with a WebSocket streaming loop, a StateStore for recent history, a rules-based health model blended 60/40 with a scikit-learn predictor, and dedicated endpoints for the assistant (/api/assistant/chat, /insight, /tts) and camera (/api/camera/turbidity).
Frontend: React 19 + Vite + TypeScript + Tailwind. Recharts for trend graphs, a custom WebGL-free camera pipeline for client-side turbidity extraction, and a slide-in AI assistant sidebar with auto-resizing chat input, voice playback, and quick-insight shortcuts.
AI stack: Cerebras for low-latency assistant responses, ElevenLabs for natural-sounding speech, with a browser speechSynthesis fallback.

Challenges we ran into

Getting consistent turbidity scores from a consumer webcam across different lighting conditions — solved by adding a calibration baseline and blending multiple features (sharpness, saturation, blue/red ratio) instead of relying on brightness alone.
Keeping the Arduino physical button and the web UI in sync for the stirrer without race conditions — fixed by having the device echo its state back on every change and letting the backend treat the device as the source of truth.
Gracefully degrading when hardware isn't connected — the backend swaps in a MockSerialReader so the whole stack still runs for demos and development.

Accomplishments that we're proud of

A full end-to-end pipeline (sensor → serial → backend → WebSocket → UI → AI) that runs live at 2 Hz without glitches.
Client-side turbidity that works on any laptop camera — no dedicated optical sensor required.
An AI assistant that actually has situational awareness of the reactor, not just a generic chatbot pasted on top.
Clean graceful-degradation story: missing camera, missing Arduino, missing API keys all fall back instead of crashing.

What we learned

How much of "industrial" monitoring is really just good UX on top of cheap sensors.
Blending a hand-written rules model with a small ML model gave us more interpretable recommendations than either alone.
WebSockets + a single central StateStore is a dramatically simpler architecture than polling everywhere, and it scales to multiple dashboards
trivially.
Prompt-caching the assistant's system context (reactor state schema + role) noticeably cut latency on repeated questions.

What's next for JERM

Closed-loop control — let the AI assistant propose parameter changes and, with user approval, actually push them to the Arduino.
Multi-reactor support with per-device dashboards and fleet-level health.
Training the ML model on real long-running culture data instead of Scripps dataset.
Mobile-friendly layout and push notifications when anomaly risk spikes.
Plug-in sensor modules (DO, CO₂, OD₆₀₀) for research-grade use.