LFM-ORBIT is a professional-product-style MVP for satellite evidence triage. It is not a production surveillance system; it is a reproducible mission-control prototype showing how local Liquid AI reasoning can reduce satellite downlink load, retain only useful evidence, and produce compact proof packets with provenance.
The product journey is simple: an operator selects a mission area, Orbit scans satellite tiles, low-value cells are pruned before downlink, retained evidence is reviewed by SAT/GND agents, and the final output is compact proof JSON with imagery provenance.
Demo guide | Validation | Release
.\run.ps1Choose 1. Install/Repair + Fetch trained Orbit GGUF -> Run.
Direct Windows command:
.\run.ps1 -InstallLinux/macOS:
./run.sh --installApp: http://127.0.0.1:5173
The default hackathon path uses SimSat/Mapbox plus bundled cached replay proof. Sentinel Hub credentials are not required.
Option 1 reuses an existing valid trained GGUF after the first download. Set LFM_ORBIT_REFRESH_MODEL=true only when you intentionally want to refresh the moving Hugging Face main handoff.
Orbit is not a single canned demo. The app ships with several reviewable mission stories so an operator can pick the signal that best explains the product loop:
| Story | What To Look For | Best Use |
|---|---|---|
| Critical Minerals Expansion Watch | evaporation pond regions, tailings regions, open-pit expansion, roads, facility clusters | clearest main showcase and provenance proof |
| Deforestation / Rondonia Frontier | canopy-loss boundary, road-edge expansion, exposed soil, retained timelapse frames | end-to-end tutorial from chat-launched mission to proof JSON |
| Fire Watch / Wildfire | burn-scar, smoke/cloud ambiguity, fireline or readiness indicators | emergency-relevance story with cautious evidence wording |
| Flood / Waterline | new surface water, overflow regions, shoreline movement | payload-reduction and visible boundary-change story |
| Maritime Activity | vessel-queue or port activity regions, link outage queueing | orbital-eclipse and compact-packet queue proof |
| Glacier / Ice-Snow | snow/ice extent, spectral-confidence guardrails, sequential timelapse context | slower science-context and abstain-safety story |
| Urban / Lifeline / Transport | road corridors, facility regions, infrastructure context | secondary operator-planning and map-context stories |
The recommended public showcase is still Critical Minerals Expansion Watch because it is visually clear and source-bound. The other stories use the same app mechanics: select an area, scan or rescan cached evidence, let SAT/GND agents review retained packets, then open Proof Mode.
- Open Mission.
- Choose Replay.
- Load Critical Minerals Expansion Watch for the shortest proof path, or choose another mission story such as Deforestation, Fire Watch, Flood, Maritime, Glacier, or Urban.
- Review Logs and Inspect for the downlinked alert, retained timelapse, source metadata, and agent notes.
- Open Agent -> Proof Mode for the compact proof JSON and visual evidence.
For judging, one complete mission is enough. The extra stories are there to show the product is a reusable mission-control prototype, not a rigid one-off recording.
cd source/frontend
npm ci
npm run demo:showcaseThe showcase loads deterministic Critical Minerals Expansion Watch replay evidence and writes video, screenshot, trace, and proof.json artifacts. No Sentinel Hub credentials are needed for the showcase path.
The primary videos are linked instead of embedded because they are larger tutorial artifacts:
- Tutorial walkthrough: plain-English product run-through from mission selection to scan, SAT/GND handoff, retained evidence, Proof Mode, compact JSON, and tagged data.
- Training journey: shows how reviewed Orbit evidence becomes reusable training data.
- Media index: all promoted videos, screenshots, story plates, and timelapse highlights.
Full repo verification:
.\run.ps1 -Verify./run.sh --verify- Tile scan over selected area and time window.
- Dual-agent triage: satellite-side prune, ground-side review.
- Retained timelapse evidence with provenance.
- Optional LiquidAI/LFM2.5-VL-450M retained-frame review when the image runtime is enabled.
- Compact proof JSON instead of raw-image downlink.
- Saved and tagged evidence for export, retagging, tuning, replay, and cached-data rescan with newer prompts or models.
The hackathon artifact is treated like a product contract: install path, deterministic demo, source-backed evidence, proof output, and honest runtime boundaries.
Region-level mining expansion evidence with provenance, target-pack context, and compact proof output.
Raw frame evidence is reduced to compact alert JSON before downlink.
Alerts queue during link loss and flush after contact returns.
Target-pack metadata travels with alerts, replay snapshots, dataset rows, and Proof Mode.
Each alert keeps provider, capture time, bbox, confidence, model metadata, and payload accounting attached.
Timelapse review uses sequential imagery slices, not static color-shift videos.
Ground Agent handles replay loads, mission packs, link simulation, and operator review cards before state changes.
Known map targets carry mission context and safe evidence guidance with the camera move.
| Check | Current State |
|---|---|
| Root verify | .\run.ps1 -Verify passing |
| Backend tests | 499 passed |
| Frontend | typecheck + build passing |
| Playwright E2E | passing with intentional skips |
| Docs/import guards | passing |
| Option 1 launch | backend 8000 and app 5173 ready |
| Clean-start smoke | idle on Atacama context, no auto replay, no default scan |
| Recorded demos | showcase, tutorial, training journey, payload, provenance, abstain, eclipse |
| Dataset export | 46 raw replay/cache samples, 34 timelapse rows |
| Retagged training set | 265 assets, 33 temporal sequences |
| Dataset | Shoozes/LFM-Orbit-SatData |
| Trained model | Shoozes/lfm2.5-450m-vl-orbit-satellite |
LFM-ORBIT uses a manifest-resolved GGUF for SAT/GND evidence-packet reasoning. Optional retained-frame image review uses LiquidAI/LFM2.5-VL-450M through the backend vision extra when enabled.
The status APIs report image_conditioned_runtime_enabled=true only after a real image adapter call succeeds.
Orbit exports reviewed evidence for retagging and tuning. The updated model handoff can be fetched back into Orbit and used to replay or rescan prior sessions.
Python 3.10+ and Node.js 20.19.0 from .nvmrc or Node.js 22.12.0+. The launchers bootstrap repo-local uv when it is not already installed.