geo.qaSee what’s happening anywhere on Earth.
geo.qa fuses public cameras, major public satellites, drones and ground sensors into one shared memory you own — then watches it, briefs you on what is actually changing, forecasts what comes next, and proves every answer with a signed receipt.
Chapter I
The agent that was never there.
Put an AI agent into production and, sooner or later, it answers a question about a place it has never observed. Fluently, and wrong.
If you have run an agent against the physical world, you have seen it. Ask whether an address flooded last week, whether a field caught rain, whether a gate stood open at three in the morning. The model reaches for the nearest static dataset and serves it back as live truth.
The failure is not reasoning. The model reasons fine. The failure is that it holds no memory of the place: nothing observed, nothing dated, nothing it can be held to.
This is a memory problem, not a model problem. So we built the memory: a record of real places, kept current by every sensor already pointed at them, that an agent can query and quote.
Chapter II
How the memory is made.
For the cameras, drones and ground sensors you run, the encoder sits at the edge of capture, so what leaves is a latent, not raw footage. For satellites, geo.qa builds on the signed embeddings of the open emem.dev ledger. Either way, a decoder turns them into citeable fact.
Observe
Sensors already in place: satellites, drones, vehicle cameras, fixed CCTV, all pointed at the same ground.
→Encode → latent
For the sensors geo.qa runs, the encoder sits at the edge of capture: a frame becomes a latent vector on-site, so raw footage never has to leave. Satellite scenes arrive already encoded, as emem.dev embeddings.
→Decode → memory
On the ground, the decoder resolves each latent onto the cell lattice: one shared, addressable Earth memory.
→File a fact
Every value is filed with its band, its time, and a signed receipt. Nothing is overwritten; the past stays queryable.
→Agent cites
Your model reads the memory and answers with the exact observations that support it, receipt attached.
Chapter III
One address for every place.
A memory is only useful if you can name a spot and trust the name. geo.qa gives every place on Earth a stable address and a simple algebra.
Chapter IV
Many modalities, one memory.
Every source writes into the same addressable record. The fleet is the unit of intelligence, not the single sensor.
| # | Modality | What it writes into the memory | Cadence |
|---|---|---|---|
| 01 | Satellite opticalmultispectral surface | land cover, water, change, vegetation indices | daily–weekly |
| 02 | Satellite SARall-weather radar | structure, moisture, flood, motion, through cloud | 1–6 days |
| 03 | Thermal / IRheat signature | flares, fires, equipment activity, anomalies | sub-daily |
| 04 | Drone surveycentimetre detail | close-range inspection of a single asset | on tasking |
| 05 | Vehicle camerasstreet level | what is round the corner, lane by lane | continuous |
| 06 | Fixed CCTVthe perimeter | gate, yard, fence line, occupancy | continuous |
| 07 | Ground sensorsSCADA / IoT | pressure, flow, level, vibration | seconds |
| 08 | Weather fieldsenvironment | rain, wind, temperature, soil moisture | hourly |
| 09 | AIS / vesselmaritime tracks | who is where on the water, and when | minutes |
| 10 | Field reportssigned human notes | a person’s observation, attributable | on event |
| 11 | Cadastral / GISthe record of record | parcels, assets, boundaries, ownership | on change |
| 12 | Acousticsound events | alarms, gunshots, machinery, leaks | continuous |
Chapter V
Where geo.qa sits.
Many are calling 2026 the year of the world model. Several layers of intelligence are converging, and each one needs something to be true about a real place.
| World & geospatial models | geo.qa | |
|---|---|---|
| Output | an inferred representation | an observed fact, dated |
| Trust | black-box inference | a signed, verifiable receipt |
| Deployment | a cloud service you call | your tenancy, airgapped |
| Time | a single snapshot | every tslot, append-only |
| Fusion | mostly one modality | many, one address |
| To build on | a fixed, pretrained model | a memory you train your own on |
World Labs’ Marble and DeepMind’s Genie generate a plausible world; they do not claim to reproduce the real one. Niantic’s Large Geospatial Model and Google’s AlphaEarth do read the real Earth, but as a model you call, not a memory you own, cite, and train on. geo.qa is the layer underneath: a dated, signed record of what was actually observed, private to you, that your own models can learn from.
Chapter VI
Train your own world model.
The memory is not only something to query. It is a substrate to learn from. Every dated, multimodal fact is supervision for a model you train, own, and deploy, without your data ever leaving the boundary.
# train a world model on your own memory (Max plan) curl -s -X POST https://geo.qa/api/world-model/training \ -H "Authorization: Bearer $GEOQA_KEY" \ -d '{ "bands": ["optical","sar_vv","thermal","weather"], "horizon": "7d" }' # → { "job_id": "wm_…", "status": "queued" }
Chapter VII
Airgapped by construction.
Because what moves is latents and signed facts, and the memory is a tenancy you own, your model can run sealed, train sealed, and still be audited.
What never leaves
the boundary holds by design, not by policy
- ◇Raw imagery. Encoded at the edge of capture; only the latent vector moves on.
- ◇Your memory. A single tenant: your keys, your retention, your jurisdiction.
- ◇Your models. Trained and served inside your boundary. No calls home, no training on your data by us.
What you keep
provenance that survives a hostile review
- ◇Provenance. Every answer carries a signed, verifiable receipt.
- ◇Audit. Replay any decision against the exact observations it used.
- ◇Interop. The same call shape as the open emem.dev protocol.
Chapter VIII
An interface agents already speak.
One call shape, shared with the open emem.dev protocol, so an agent reads public and private memory in a single answer.
# ask the live API about a place — land cover from satellite curl -s \ "https://geo.qa/api/tool/satellite/lulc/analyze?lat=25.276&lon=55.296" \ -H "Authorization: Bearer $GEOQA_KEY" # → { "class": "water", # "coverage": { "water": 0.94, … }, # "source": "sentinel-2" }
Chapter IX
Questions from the field.
The same memory answers very different questions. A few of the surfaces geo.qa is built for.
Autonomous fleets
The vehicle’s cameras write to the memory; the car reads neighbouring cells to know what is round the corner before it sees it.
what is in cell 8a2a1072b… right now?
Energy & infrastructure
Pads, lines and stations watched by every sensor at once. One page goes out only when independent witnesses agree.
did pressure at G-247 drop overnight?
Insurance & risk
Price and settle against what was observed, not what was claimed, and every fact arrives with its receipt.
was this parcel underwater on 04 May?
Defence & GEOINT
An owned, airgapped memory with full provenance: answers built to survive an audit, weeks later.
what changed at this site since Tuesday?
Chapter X
The world, read back to you.
Most platforms stop at storage. geo.qa reads its own memory: it briefs you on what is changing, forecasts what comes next, and proves every word — grounded only in what was actually observed.
It briefs you
Live World Intelligence synthesises what is happening across your shared memory — a situational briefing built only from signed observations, never invented. Quiet places are reported as quiet.
what is happening across my sites right now?
It forecasts — with proof
The next‑step state of any cell, with an honest skill‑vs‑persistence score and a signed receipt. It tells you plainly when it cannot beat the baseline, rather than dressing up a guess.
what comes next here — and can I trust it?
It proves every answer
Every fact carries its cell address, its source, and a cryptographic receipt you can verify offline against emem.dev. The proof travels with the answer.
show me the evidence.
Give your model a memory.
Stop letting it guess about the world. Start letting it cite, and train on what it cites.
geo.qa · a vortx ground decoder · emem.dev open protocol