Most software is written on the assumption that the network is present and the server is authoritative. The client asks, the server answers, the screen updates. When connectivity drops, the application degrades into a spinner and waits for the world to come back.
A great deal of the work we do cannot make that assumption. A conservation gate in the Ngorongoro landscape, a screening room in a rural health facility, a sensor plot in a research field — these are places where connectivity is intermittent by nature, not by fault. Treating the outage as an error state means the system is broken most of the time.
The assumption inverts
Offline-first inverts the relationship. The local device is authoritative for what it has observed. The server is a place where observations from many devices are eventually reconciled. Connectivity becomes an opportunity to synchronise rather than a precondition for operating.
Stated that way it sounds like a small change. In practice it reaches into almost every design decision.
- Identifiers must be generatable on the device, because a record cannot wait for a server to name it.
- Writes must be expressible as intents that can be replayed, not as final states that overwrite.
- Conflict resolution has to be a deliberate, domain-specific decision — last-write-wins quietly destroys data.
- The interface has to tell the user what is captured, what is pending and what is confirmed, without making them think about synchronisation.
- Storage on the device must be bounded, because the sync window may be days away.
Our systems are often built for places where connectivity, power and infrastructure can’t be taken for granted. That reality is a design input from day one.
Power is the constraint behind the constraint
Connectivity gets the attention, but power is often the harder limit. A device that must run on solar with a battery buffer has a duty cycle, and that duty cycle sets the ceiling on everything else — how often a camera can capture, how much local inference is affordable, how long a radio can stay awake.
This is one reason edge inference is attractive in these deployments even when a cloud model would be more accurate. Sending a full image over an expensive, intermittent link costs far more energy than running a smaller model locally and sending a result. The right question is rarely "which model is most accurate" but "what is the smallest thing I can send that still carries the decision".

Degradation should be designed, not discovered
Every system in these environments will operate in a degraded mode at some point. The difference between a robust deployment and a fragile one is whether those modes were designed deliberately or arrived at by accident.
We try to write down the degradation ladder explicitly: full connectivity, intermittent, local-only, and device-unavailable. For each rung, what still works, what queues, what is refused outright, and what the operator sees. A gate that cannot reach the network should still admit a vehicle with a valid permit it has already cached — and it should say so plainly, rather than failing in a way that looks like the permit is invalid.
