Why us
Own the system. Own the data. Keep it here.
Most survey products are a black box that takes a council's footage away, processes it somewhere unstated, and hands back a report. We built the opposite: our own models, trained in the open on data we're entitled to use, running on Australian infrastructure, on footage that has already had every face and number plate removed before it left the phone.
At a glance
| Velospatial | Common practice | |
|---|---|---|
| Who owns the survey data | The council. Footage, defect registers and evidence packs belong to the council that created them, exportable in full at any time. Leaving takes your data with you. | Data is frequently held in the vendor's platform under the vendor's terms, with export limited, paywalled, or lost on exit. |
| Where it lives | Google Cloud Australia: imagery and processing in Sydney, the database in Melbourne, and we are working through the remaining steps so the data always stays here. | Often unstated, or offshore by default because that is where the vendor's platform happens to run. |
| Faces and number plates | Detected and irreversibly pixellated on the phone, before upload. What leaves the device never carries an identifiable person or plate. | Commonly blurred server-side after upload, which means identifiable footage travels and is stored before redaction, or redaction is left to the council. |
| Who owns the models | We do. Built and trained in-house on openly licensed imagery and our own labelled corpus, with attribution shipped and copyleft licences excluded from day one. | Licensed third-party models, or models trained on data of unstated provenance that the vendor could not defend line by line. |
| How the AI improves | Every disputed reading becomes a human-labelled training example, so the models learn Australian roads, this council's roads, from the corrections people make. | A fixed model that improves on the vendor's schedule, trained on other countries' roads. |
| When the AI is unsure | Low-confidence frames route to a person. No model publishes a verdict on its own. | Varies; confidence handling is often invisible to the customer. |
The detail behind each claim
Survey imagery and processing run in Google Cloud's Sydney region and the database in Melbourne. Data residency isn't a tick-box for a council, it's a procurement condition, and ours is verifiable, not asserted.
Face and plate blurring runs on the device, in the capture pipeline itself. The blur is pixellation, not an overlay: the information is destroyed, not hidden. Our own plate detector plus a text-shape fallback means a missed plate has a second net.
The footage a council captures and the registers built from it are the council's records. Full CSV and evidence-pack export ship with the product, because data you cannot leave with is data you do not own.
Every training source is Public Domain, MIT or CC BY, with attribution shipped and licences re-checked automatically at fetch time. We can show a council, or a lawyer, where every training image came from. The technology page shows the numbers.
The phone's model reads every frame and is marked against a stronger model; every disagreement is labelled by a person and feeds the next training run. The system gets better at Australian roads with every drive, and that improvement stays ours.
Models propose, humans dispose. Defects the AI is unsure about, disagreements between models, and anything a council disputes all route to a person, and the review record is part of the product.
The honest print. Storage and processing already run in Australia, and we are working through the remaining steps to ensure the data always stays here: one AI adjudication step currently uses a global cloud endpoint, and it moves onshore the day the provider offers that model family in an Australian region. Every frame that step sees has already been de-identified on the device.
Comparisons opposite describe common patterns in the market, not any single vendor. Where a competitor does better on one of these, we would rather you ask them the question than take our word for it.
Ask us the hard questions
Data residency, licensing, model provenance, privacy: the answers are written down, and we'll show you the evidence behind any row of that table.