Liquid Lens

On-board Water Quality Monitoring

Satellites generate terabytes of multi-spectral imagery per pass, but a small-sat's downlink budget is limited. Raw imagery cannot fit. Under the traditional approach, data queues until a ground station appears, then waits again for inferencing and analyst review - a process that can take hours or days. By the time an algal bloom or turbidity spike reaches a decision-maker, it has already spread.

Liquid Lens moves the decision on-board. A two-tier inference pipeline runs directly on the satellite:

• Tier 1 - Spectral Math: Four spectral indices (NDWI for water presence, NDTI for turbidity, NDCI for algae, NIR absorption for depth) classify every observation, with no GPU required. The majority of observations are normal and require no downlink traffic.
• Tier 2 - Vision Model (conditional): Only when Tier 1 flags a medium or high risk anomaly does Liquid AI's LFM2.5-VL-450M activate. It visually confirms the spectral finding and generates a natural-language explanation - turbidity plumes, algae mats, exposed lake beds - ready to transmit as a compact alert. The result is global lake monitoring at small-sat cost, with actionable alerts produced in seconds from the moment the satellite is overhead.

SimSat API (:9005)
      │  multi-spectral bands: green · red · NIR · red-edge
      ▼
 collect_data.py
  ├─ Tier 1: Spectral indices (NDWI / NDTI / NDCI / NIR)
  │          no GPU required · filters normal passes
  │
  └─ Tier 2: LFM2.5-VL-450M via llama-server (:8080)
             triggered only on high / medium risk
             produces natural-language anomaly description
      │
      ▼
 shared_state.json  ──►  Flask app (:5000)  ──►  Plotly globe dashboard
                         live satellite position · risk level · VLM insight  

• Docker & Docker Compose (v2)
• wget, unzip (for setup script)
• ~500 MB free disk space for model weights (downloaded by setup script)
• Additional disk space in data/ grows with simulation time as captured spectral images are saved per lake per pass
• SimSat (DPhi Space simulator) — must be running before starting Liquid Lens (see step 1 below)

git clone https://github.com/DPhi-Space/SimSat
cd SimSat
docker-compose up

Leave this running.

• Dashboard: http://localhost:8000
• API: http://localhost:9005

git clone https://github.com/LiquidLensSystems/water-vlm
cd water-vlm
chmod +x setup.sh
./setup.sh

This downloads the LFM2.5-VL model weights (~500 MB) and the Natural Earth lakes shapefile into models/ and data/ respectively.

llama-server runs directly on the host so it can access CPU/GPU backends. Open a dedicated terminal and leave it running:

cd llama-b7633
./llama-server \
  -m ../models/LFM2.5-VL-450M-Q4_0.gguf \
  --mmproj ../models/mmproj-LFM2.5-VL-450m-F16.gguf \
  -c 8192 \
  --port 8080 \
  -ngl 99

Use -ngl 0 if you do not have a GPU. Wait for llama server listening before proceeding.

docker-compose up --build

• Dashboard: http://localhost:5000
• llama-server: http://localhost:8080

water-vlm/
├── event_radar/
│   ├── collect_data.py          # Satellite loop + tiered inference
│   └── mission_control/
│       ├── app.py               # Flask dashboard backend
│       └── templates/
│           └── index.html       # Plotly globe frontend
├── llama-b7633/                 # llama.cpp b7633 binaries (run on host)
├── docs/
│   └── dashboard.png            # Dashboard screenshot
├── models/                      # Downloaded by setup.sh (git-ignored)
├── data/                        # Shapefile + captured images (git-ignored)
├── docker-compose.yml
├── Dockerfile.app
├── requirements.txt
└── setup.sh

• network_mode: host is used so all services can reach SimSat at localhost:9005 and llama-server at localhost:8080 without extra network configuration. This works on Linux and WSL2. It does not work on macOS Docker Desktop.
• llama-server runs on the host directly to ensure CPU/GPU backends load correctly.
• Model weights and captured data are git-ignored. Run setup.sh before first use.
• Always use docker compose down -v (not just docker compose down) when restarting after frontend changes, to ensure the volume is refreshed.