instead of buying another raspberry pi or letting an old xiaomi redmi 9i rot in a drawer, i turned it into a 24/7 edge ai node and local s3/sqlite storage server running on about 1.2w idle and 4.2w peak.
hardware:
mediatek helio g25 (8x cortex-a53 cores)
4gb lpddr4x ram
emmc 5.1 internal flash + microsd
built-in 5000mah battery acting as a zero-cost ups
the software stack:
rooted android base running a linux chroot with partial wakelocks so android's aggressive lowmemorykiller daemon and doze mode don't quietly murder background processes.
speech-to-text: whisper.cpp (base.en) compiled with arm neon simd vector instructions on the a53 cores.
llm: llama.cpp running quantized qwen 2.5 (0.5b instruct q4_k_m).
memory governor: an in-process jit governor that automatically evicts idle models back to disk after 75 seconds so the 4gb ram never gets oom-killed.
storage & db: project-isolated sqlite in wal mode plus an s3-compatible blob store with an in-memory hot ram index.
the real headache: grid outages & zombie tunnels
the power cuts out roughly once a day here. the phone’s battery means the node itself never shuts down, but downstream chaos was the real killer: the router dies, wi-fi drops, and temporary cloudflare quick tunnels would instantly get killed with http 530s or freeze on stale tcp sockets once the connection returned.
to make it genuinely self-healing, i moved to a named tunnel with a persistent domain and wrote a background supervisor watchdog. the second wi-fi comes back up post-outage, it flushes dead socket states, tears down zombie listeners, and resyncs the cloudflare edge worker so the database, storage, and inference endpoints recover without me ever touching the phone.
actual numbers:
qwen 2.5: ~7 tok/sec
local sqlite / s3: sub-millisecond on lan
thermals: sits around 32-34°c with no active cooling
Top comments (1)
The battery acting as a zero-cost UPS is great until the charging circuit swells the pouch after six months pinned at 10%. If the kernel exposes charging control via sysfs or an ACC daemon, capping the charge limit at 7% saves the hardware from turning into a spicy pillow while keeping enough buffer for grid drops.