The Android layer for unattended industrial devices

Your devices are where
nobody is. That is where the engineering starts.

An unattended device cannot wait for a site visit. It has to stay alive on its own, take updates remotely, and recover by itself when something breaks. None of that is solved in the app — it lives in the Android platform underneath. That layer, on Qualcomm hardware, is what we do.

COREEDGE LAB / DEVICE WORKBENCH ANDROID · IOT · EDGE AI
COREEDGE ON TOP Product App end customer Device Owner board vendor · cloud IO SDK / SERVICE API SYSTEM UID CoreEdge Android Layer BSP · UPGRADE · NETWORK PATH · MEMORY · BOOT ANDROID BSP / HAL / KERNEL SoC · drivers · SELinux · route SoC / BOARD industrial hardware ETH LTE Unattended site remote maintenance VPN reachable OTA staged diag bundle failover policy ONLINE VPN OTA COREEDGE THE LAYER THAT KEEPS IT RUNNING WITH NOBODY THERE
START The board boots
COREEDGE IO SDK · network path · version upgrade · memory · boot time
RUNS UNATTENDED Ships, reconnects, updates, and can be repaired remotely

Your app on top, device management around it — the layer in between is ours

The end customer maintains one app. The board vendor brings the hardware, the device owner app and the cloud. Neither of them owns the Android layer in between — the SDK that lets the app reach I/O without platform work, the logic that decides which network path traffic takes when several coexist, the version upgrade when the vendor BSP stops moving, the memory and boot time an unattended device lives or dies by. That layer is what we build.

End customer Product App

The product logic, the UI, the domain knowledge — their IP, and the reason the device exists. One app, which should not have to be rewritten because the board changed.

Board maker / integrator Device Owner · cloud · OTA

Board vendors bring their own device owner app, cloud service and update pipeline — that is their product and their recurring revenue. We do not sell a management platform and do not compete here. We make sure the Android underneath can actually be enrolled, locked down and updated by theirs.

CoreEdge Lab Android system customization
IO SDK Network path selection Version upgrade BSP Memory Boot time Self-recovery Edge AI
Silicon vendor / hardware Chip and firmware

Modem, ISP and RF calibration, signing keys, certification. Not our layer, and we say so up front — work that needs those things cannot be shipped without them.

Build it before it ships. Keep it alive after.

Most teams get stuck in the first half, then discover in the second half that there is no way back — the device has shipped, nobody is on site, and the update path was never designed in. These two halves have to be planned together; the second one cannot be bolted on from a distance.

01

Before it ships — making the device exist

Board bring-up, driver porting, HAL work, Android version upgrades, memory and boot-time reduction, edge-AI on the NPU. This is the part the SoC vendor and the board maker both leave to you.

Bring-upDriverHALVersion upgradeSystem slimmingEdge AI
02

After it ships — keeping it alive with nobody there

OTA releases, security patches, remote debugging over a private tunnel, fleet visibility, and self-recovery when a device wedges. If this was not designed in before shipment, there is no way to add it afterwards from a distance.

OTASecurity patchesRemote debugFleetSelf-recoveryA/B rollback

What we do not take on

One rule: anything that requires the silicon vendor’s signing keys, licensed tools, or proprietary calibration data, we decline. Not because it is hard — because taking it on without those things produces work you cannot ship.

We decline

  • MIPI camera ISP / 3A tuning
  • Battery charging curves, fuel gauge, safety certification
  • Modem / aDSP / SLPI firmware and calibration
  • RF calibration
  • CTS / GTS and regulatory certification
  • Widevine L1
  • Production signing key custody
  • Hardware design and component selection

Need Google Play and GMS? That requires a Google agreement plus per-model CTS/GTS approval. For those products we would point you at hardware that already carries the certification, rather than take the work.

These look adjacent — we do take them

  • UVC / USB cameras and hot-plug recovery
  • Vision AI pipelines end to end
  • Inference on the on-chip NPU (public SDK)
  • DSP fault diagnosis and configuration-layer fixes
  • Power and thermal problem localization
  • Audio codec, amplifier and routing bring-up

The line is not "hardware vs software". It is whether the work can be completed and verified without something only the vendor can give you.

Three Ways to Work With Us

Each engagement has a defined scope, deliverable, and acceptance — so you know exactly what you are buying before we start. Deep focus on the Qualcomm Android platform.

Stop the bleeding

Rescue Sprint

Project quote · 4–8 weeks · defined acceptance

For bring-up that is stuck, a platform unstable in the field, or integration that has stalled.

  • Scoped fix within a fixed window
  • Stabilized, repeatable system baseline
  • Acceptance criteria and handover docs
After shipment

Platform Maintenance Retainer

Monthly · scaled by fleet size / SLA / release cadence

For devices already in the field that need ongoing ownership of platform availability.

  • OTA release review and rollback readiness
  • Diagnostics and remote recovery
  • Field issue tracking and vendor coordination

Not sure which fits? Start with a Paid Assessment — it credits toward a follow-on engagement.

Every number below was measured on hardware

Client work, anonymized. Same device, same conditions, before and after — no ranges, no "up to", no "varies by scope". Where we do not have a measured number, there is no case here.

CONNECTIVITY

LTE that stopped dropping

A field device was changing IP roughly every 24 seconds and had no network at all for 16% of the time. The recorded cause was "modem power saving." It was not. The allowed radio-access list was too wide, so at the edge of coverage the modem kept retrying 3G and 2G, tearing the connection down each time.

Same device, same SIM, same location: IP changes 40 → 0 over 15.8 minutes; time without network 16% → 0%. Signal was weaker in the fixed run than in the broken one.
RAT policyallowedNetworkTypesrmnetPDN
MEMORY

The camera app the system kept killing

A 3 GB unattended device is healthy when idle — the risk only shows up under camera load, when buffer allocations spike and the kernel starts killing by score. It killed the customer camera app and left the process that caused the spike running. Per-app memory limits could not catch it: the kernel accounting they use does not include those buffers at all. So the fix was a watchdog that stops the actual culprit instead of waiting for the kernel to pick a victim.

Same stress test, same device: available memory bottomed out at 334 MB without the watchdog and recovered to 1823 MB with it, in 2–4 seconds. The camera app kept the same PID throughout — it was never killed. Separately, background resident memory went from 496 MB to 227 MB on the same image.
ION / dma-bufmemcglmkdoom_score_adj
VENDOR STOPPED UPDATING

When the vendor stops updating, the system layer is still yours

The board was stuck on Android 13, the vendor BSP was no longer maintained, and there was no source to work from — every problem became a wait for a release that was never coming. We replaced only the system image and left the vendor partition untouched: not one byte changed. Getting to a newer Android was the side effect; the point is that the system layer is yours to change again.

Android 13 → 16 on the same hardware. 40 vendor HALs, GPU, Wi-Fi 6, 28 sensors and cellular all kept working. Telephony verified end to end with 0% packet loss.
GSITrebleVINTFVNDKAVB
BOOT TIME

Cutting boot time by more than half

An unattended device that takes 34 seconds to come back after a power cut is 34 seconds of missing service, every time. Most of it turned out to be the kernel writing every log line to a serial console nobody was reading.

Boot time 33.7s → 14.5s on the same hardware, verified on device.
kernel cmdlineinitbootchart
EDGE AI

Object detection from 140 ms to 3 ms

A detection model ran on the CPU because moving it to the NPU made it slower, not faster. Offline per-operator profiling found a single pooling layer running 250× slower than expected on that accelerator; replacing it with an equivalent convolution unblocked the whole graph.

Inference latency 140 ms → 3 ms on the on-chip NPU. Lesson: when the accelerator is slow, profile it — do not guess.
NPUQNNper-op profilingYOLO
SYSTEM SLIMMING

Making the OS fit the board you already have

The customer board had 2082 MB of usable space and the stock image needed more. Rather than repartition every unit in the field, we removed what an unattended industrial device never uses — consumer apps, unused runtime versions, virtualization support.

System image 1982 MB → 1684 MB. Headroom on the customer board went from 82 MB to 397 MB, and no partition change was needed.
SoongPRODUCT_PACKAGESVNDKsuper

Technical Notes From the Problems We Actually See

These articles explain the engineering gaps behind industrial Android devices: BSP customization, network recovery, OTA, diagnostics, and long-term platform maintenance.

Who does the work

James Lee, Founder and Principal Engineer of CoreEdge Lab
James Lee
Founder & Principal Engineer

CoreEdge Lab is led by James Lee, with 17+ years across embedded systems, IC design support, Android BSP / SoC platform work, and production delivery. Clients usually come to us when the platform is already behind schedule, unstable in the field, or missing the system-layer ownership needed to ship. We work where those problems are actually solved: bring-up, integration, recovery, diagnostics, and production hardening.

17+ Years in Embedded
IC → BSP IC Design to Android System
E2E Boot to Production

Book an Initial Assessment

Tell us where the platform is stuck: bring-up, unstable networking, OTA, diagnostics, or production readiness. We'll reply within 24 hours with the most likely next step.

Best fit: ODM/OEM device teams, industrial Android products, Edge AI boxes, rugged handhelds, and SI teams that need platform depth without building a full BSP team.
More context (optional)

By submitting, you agree to our Privacy Policy. We only use your info to respond to your inquiry.

Prefer email directly? [email protected]