Offline classification
Five task families, no connection.
Technology · Inference · Whitepaper
Escalate only when it helps.
Inference is arranged as a ladder: on-device classifiers, then a bundled corpus, then retrieval, then a frontier model — and each rung is only climbed when a link exists and the escalation would measurably improve the answer. Most queries never leave the handset.
01 / Problem statement
What this has to solve.
Two failure modes dominate AI features in field software.
Everything goes to the cloud. The feature is unavailable exactly when it is needed, and every trivial classification costs a request. Users learn not to rely on it.
Nothing goes to the cloud. A small local model is asked open-ended questions it cannot answer and produces confident nonsense, which is worse than no answer in a safety context.
The correct arrangement is a router that knows which rung a question belongs on, and an interface that is honest about which rung answered it.
02 / Mechanism
The design, component by component.
- On-device classifiers
- TensorFlow Lite models for wound assessment, plant and wildlife identification, celestial recognition and audio classification. Quantised for handset inference, shipped with the app.
- Bundled knowledge base
- A curated wilderness and medical corpus on the device, so a grounded answer with a citation is possible with no link at all.
- Model router
- Each query is scored for the capability it requires. The router selects the cheapest rung that can answer it and records which rung answered.
- Connectivity gate
- Escalation requires both a link and an expected improvement. A question the local corpus answers well is not escalated merely because a link happens to exist.
- Retrieval
- Vector retrieval over the full corpus when online, returning sources rather than an unattributed generation.
- Confidence surfaced
- Classification returns a distribution, not a verdict. The interface shows confidence bars, and a low-confidence result is presented as such.
- Suggested actions
- An answer may propose a concrete action, which a human authorizes. The model never acts.
- Jetson-class target
- For deployments requiring more capable local inference, the target is Jetson-class edge hardware with a distillation pipeline for the models that need it.
03 / Properties
What this guarantees.
Grounded offline answers
From a bundled, cited corpus.
Cheapest-rung routing
With the rung recorded and shown.
Honest confidence
Distributions, not verdicts.
Never metered on-device
Local inference costs us nothing, so it costs you nothing.
Human-authorized actions
The model proposes; a person decides.
Edge hardware target
Jetson-class for deployments that need it.
04 / Reference
| Rung | Runs on | Needs link | Used for |
|---|---|---|---|
| TFLite classifier | Handset | No | Images, audio, sky |
| Bundled corpus | Handset | No | App and wilderness questions |
| Local reasoning | Handset | No | App-level queries |
| Retrieval | Server | Yes | Full-corpus grounded answers |
| Frontier model | Server | Yes | Genuinely open-ended questions |
| Edge accelerator | Jetson-class node | No | Deployments needing more locally |
05 / Process
Traced, node by node.
Most questions never leave the phone.
On-device work is unmetered on every paid tier. Only escalation counts.