+++ title = “Acme Camera” description = “” date = 2026-10-13

[extra] reveal_theme = “black” reveal_transition = “slide” +++

Acme Camera

A Proposal

Agenda

Architecture: hardware and software
Performance metrics and targets
Instrumentation and fleet management
Risks, plan and questions

Note: Roughly 15 minutes each, questions welcome throughout Detail for every software component is in the appendix if we want to dig in

Product Overview

A wireless security system architecture supporting long operational life

Note: The pitch: months on a charge, a person-confirmed alert in seconds, and fleet data that proves both

–

Requirements

Battery powered, months per charge
Outdoor: weatherproof, −20 to 50 °C
Person-confirmed alerts in seconds
Night vision and a spotlight
Setup from a phone, Wi-Fi to the home
Optional solar

Note: Matches the example products: a standalone battery camera and a solar-topped variant Temperature range matters for the battery more than anything else, cold eats usable capacity

Assumptions

Area	Assumption
Battery	~6000 mAh Li-ion pack
External power	regulated 5 V, solar included
Activity	~20 events/day, ~10 s clips
Network	2.4 GHz Wi-Fi to a home AP
Recording	event clips, live view on demand
Backend	exists, we own the device side

Note: These drive every number later in the deck, so happy to rerun the math if Acme’s are different Solar is just another 5 V source, the panel’s own controller handles MPPT No continuous recording, a battery camera can’t afford it

Hardware Design

Note: Two domains: an always-on MCU with the sensors, lights, PMIC and BLE / Thread radio, and a mostly-asleep SOC with the camera, storage and Wi-Fi The MCU decides when the SOC is worth waking

SOC + MCU

–

MCU

ultra-low power attached to motion sensors, lighting, power control, and the BLE / Thread radio

SOC

attached to camera/storage/Wi-Fi and spends most of its time in deep sleep

–

MCU Candidates

nRF54L: BLE 6, Thread / Matter
nRF52840: BLE 5, Thread / Matter

SOC Candidates

Rockchip RV1106
Ingenic T41
Ambarella CV7

Note: Radio split: Wi-Fi lives on the SOC for video, BLE and Thread live on the MCU None of the SOCs have Wi-Fi on-die, so it’s an SDIO module next to the SOC BLE on the MCU handles setup from the phone app without waking a soft-AP Thread / Matter lets the camera show up as a motion sensor or light in a smart home while the SOC sleeps Neither MCU has Wi-Fi; the nRF54L can host an nRF7002, but putting video through the MCU is the wrong path

Motion

Configurable

PIR

plus

MMWave

Note: Power Efficient Setup, PIR is used to wake the SOC and then mmwave is used to determine if people are present

LED Driver

to support Infrared and Visible Spots for lighting areas for improved visibility

Environmental

Temperature
Luminosity

Power Subsystem

Battery Charge Controller & Gauge

–

What about solar?

Let’s set the expectation that any input power source provides regulated 5V, solar included

–

Solar Charge Controller

TI bq25504
e-peas AEM10941 / AEM13920:

Note: nPM1300 does not provide the MPP support for the solar cells so we need an external IC to properly do it

Software Design

Note: mcuctl sets the motion and light mode on the embedded mcu powerctl controls the power domain for the battery charger and reads the fuel gauge portalctl provides the initial configuration UI and subsequent status web portals cameractl reads frames and dispatches them to whatever has been configured configctl is the source of truth that gets configuration requests broker: mqtt, likely mosquitto infctl: inference controller for on-chip npus to load models (bonus)

Build & OS

Yocto on the SOC vendor’s BSP
Read-only rootfs in A/B slots, writable /data
Signed images, built reproducibly in CI
Rust daemons as sandboxed systemd units

Note: Yocto for layers, an SDK for app developers, and license manifests / SBOM for free; Buildroot if the chosen vendor only supports that Kernel trimmed to what the board needs, suspend-to-RAM tuned so a wake is a resume, not a boot Read-only rootfs means an update is a whole-slot swap, nothing drifts between devices /data holds config, credentials and the clip buffer, and is the only thing a factory reset wipes

–

Process Model

Unit	Can reach
broker	`/run/broker.sock`
mcuctl	UART to the MCU
powerctl	PMIC over I²C
cameractl	sensor, ISP, clip buffer
portalctl	BLE via mcuctl, HTTP
configctl	`/data/config`
telemetryctl	`/data/telemetry`, uplink

Note: Each unit runs as its own user with ProtectSystem=strict and DeviceAllow for only its hardware The broker checks the connecting uid (SO_PEERCRED) against per-topic ACLs, so portalctl can request config but never write it A crashed daemon is restarted by systemd and counted in telemetry

–

Boundaries

Boundary	Transport
process ↔ process	MQTT on a Unix socket
SOC ↔ MCU	UART + wake GPIO
device ↔ cloud	MQTT 5 over mutual TLS
clips → cloud	HTTPS to object storage

Note: The MCU raises a GPIO to wake the SOC, then they talk COBS-framed messages over UART from one shared schema, so both sides agree on MotionDetectionMode and friends mcuctl is the only process that touches the UART, everything else goes through the broker Video never goes over MQTT, cameractl gets a pre-signed URL and uploads clips directly

Boot Plan

Note: Uses an A+B boot that only changes the boot target once we’ve booted successfully and connected to the backend

Security

Trust from boot to cloud

Note: Home security buyers ask about this first, and a compromised camera is a camera pointed into someone’s home

–

Layers

Layer	Measure
Boot	secure boot, signed A/B images
Keys	secure element / OTP, never in rootfs
Storage	encrypted `/data` and clips
Processes	per-service users, sandboxed units
Broker	per-topic ACLs by peer uid
Setup	BLE pairing, soft-AP as fallback
Cloud	per-device certs, mutual TLS

Note: The claim token from setup is exchanged for a per-device cert, nothing shared across the fleet Telemetry carries metrics only, no video and nothing personal beyond the display name Factory reset wipes /data and revokes the device cert server side

Device ↔ Cloud

One connection, held only while awake

Note: The device is asleep most of the time, so the cloud is designed around a device that is usually unreachable

–

Connection

Concern	Approach
Identity	per-device cert from the claim
Transport	MQTT 5 over mutual TLS
Reconnects	TLS session resumption
Commands	queued until the next wake
Config	device shadow, desired vs reported
Offline	buffer locally, send on reconnect

Note: Session resumption skips most of the TLS handshake on every wake, which is time and energy on the event path A persistent MQTT session holds commands for a sleeping device, it picks them up on its next wake or heartbeat Live view needs the device reachable while asleep, which means Wi-Fi power-save on the SOC module or the keep-alive radio from the trade-offs, both cost standby current

–

Cloud Topics

Topic	Direction	Carries
`d/<id>/event`	up	motion, detections
`d/<id>/telemetry`	up	batched metrics
`d/<id>/shadow`	both	desired / reported config
`d/<id>/cmd`	down	live view, reboot, logs
`d/<id>/ota`	down	update offers

Note: The device can only publish and subscribe under its own id, enforced by the cert The shadow is the bridge to configctl: the cloud writes desired, configctl applies it and reports back what it actually runs

Components

Component	Role
broker	MQTT on a Unix socket
mcuctl	motion, lights, BLE via the MCU
powerctl	power modes, PMIC telemetry
portalctl	setup and status portals
cameractl	frames to storage and cloud
configctl	source of truth for settings
infctl	on-device inference (bonus)
telemetryctl	buffers and uploads metrics

Note: Details and types for each are in the appendix, the links jump there

Anatomy of an Event

Step	t
PIR trips, MCU wakes the SOC	0
SOC resumes from suspend	~200 ms
Radar confirms a person	~300 ms
First frame from the sensor	~700 ms
Wi-Fi reconnects	~1.5 s
Notification sent	~2 s
Clip uploaded, back to sleep	~15 s

Note: The SOC resumes in parallel with the radar check, if the radar doesn’t confirm it goes straight back to sleep Capture starts before the network is up, frames go to the local clip buffer first, so the first frame isn’t waiting on Wi-Fi ~15 s at ~400 mA is the ~1.5 mAh per event the power budget uses Keeping Wi-Fi associated while asleep would cut 1.5 s but costs standby current, a config trade-off

Metrics

Performance within the configured power envelope

leading to insights for better defaults

Battery life across the fleet

Note: Every config choice (motion mode, light mode, schedules) trades responsiveness for battery, so the metrics are about how well a device does inside the envelope its owner picked, and fleet data tells us which defaults to ship

–

Targets

Metric	Target
Battery life, default config	≥ 3 months
Standby draw, SOC asleep	≤ 100 µA
Motion → first frame	< 1 s p95
Motion → notification	< 5 s p95
False wakes (no person)	< 20%
Crash-free wakes	> 99.9%
OTA success, first try	> 99.5%

Note: Assumptions: ~6000 mAh pack, 20 events/day, 10 s clips, ~400 mA while the SOC, Wi-Fi and IR are up Each event costs roughly 1.5 mAh, so events dominate and standby has to stay tiny Wake latency is the user-visible one: a late first frame means a clip of someone’s back False wakes are PIR triggers the radar doesn’t confirm, they cost battery for nothing A/B boot means a failed OTA rolls back rather than bricks, the target is about not needing to

–

Power Budget

Item	mAh/day
Standby, 100 µA × 24 h	2.4
Events, 20 × 1.5 mAh	30
Heartbeat + telemetry	~0.5
Self-discharge, ~2%/month	~4
Total	~37
6000 mAh × 80% usable	~130 days

Note: Events are ~80% of the budget, so the false wake target and clip length matter more than standby 40 events/day halves it to ~65 days, cold weather takes another chunk off usable capacity This is the table the fleet data should confirm or correct

–

Slicing the Fleet

Firmware version
Motion / light / power mode
Battery vs solar vs external
Temperature and Wi-Fi signal

Note: Averages hide problems: a battery life drop on one firmware, or in cold climates, or on weak Wi-Fi These are the dimensions every metric gets broken down by Firmware cohorts gate staged OTA rollouts: a regression in one cohort halts the rollout

Instrumentation

What we measure, how it leaves the device, and how often

Note: Combines environmental and fuel gauge data with what each component already knows

–

Signals

Source	Signals
powerctl	voltage, current, charge
mcuctl	wake reason, radar confirm
cameractl	wake → frame latency
network	connect time, RSSI
system	fw, boot slot, crashes
environment	temperature, light

Note: powerctl also reports time to empty, mcuctl how long the lights were on, cameractl clip length, network upload time, system the reboot reason Each component publishes its own metrics to the broker, so nothing new has to reach into another process Energy per wake comes from the fuel gauge before and after, which ties every event to its battery cost

–

Collection

Components publish metrics to the broker
telemetryctl aggregates and buffers to flash
MCU logs hourly samples while the SOC sleeps
Uploads ride along when Wi-Fi is already up

Note: Never wake the radio just to report, telemetry should be well under 1% of the energy budget The MCU keeps a small ring buffer of sleep samples (battery, temperature, wake counts) that the SOC drains on its next wake Buffered to flash so a dead battery or lost network doesn’t lose history Upload is MQTT over TLS to the backend, the same connection the device already holds for events

–

Reporting Cadence

When	What
Every event	wake summary: reason, latency, energy
Daily heartbeat	counters, battery state, sleep samples
After a crash or rollback	coredump and reboot reason
On demand	debug logs, per device

Note: The heartbeat goes out even with no events, so a silent device is a signal on its own Coredumps are symbolicated server side, the device only ships the raw dump On demand logs are for support cases, off by default to save power and bandwidth

–

Fleet Dashboards

Projected battery life per device vs target
Latency and false wake rate per cohort
Alerts: projected life below target, missed heartbeat, crash spikes after OTA

Note: Projected life comes from the measured drain rate, so we can flag a device months before it dies Missed heartbeat for 48 h means dead, offline or stuck, worth an owner notification A crash spike in a new firmware cohort pauses the rollout automatically

Fleet Management

Identity, config and updates for every device

Note: Two audiences use the same tools: Acme operating the whole fleet, and owners managing their own cameras

–

Lifecycle

Stage	What happens
Factory	keys in the secure element, cert registered
Claim	owner’s app links the device to an account
Operate	shadow config, telemetry, alerts
Update	staged OTA by cohort
Retire	reset wipes `/data`, cert revoked

Note: The claim token is the QR code on the device, it’s useless without the device cert behind it A resold camera goes through Retire then Claim, so the new owner never sees the old owner’s data

–

Bulk Configuration

Owners group cameras by site or area
Profiles: motion, light, power, processing
Apply a profile to a group, override per camera
Each camera reports when it has applied

Note: A profile write sets the desired shadow on every camera in the group, each one picks it up on its next wake or heartbeat configctl validates it like any other change, so a profile can’t push a camera into a state it doesn’t support The app shows pending / applied / rejected per camera, because sleeping devices apply changes over minutes, not instantly Same mechanism Acme uses at fleet scale, e.g. changing a default for a firmware cohort

–

OTA Rollout

Stage	Devices	Gate
Internal	~50 dogfood	a clean week
Canary	1%	crashes, drain vs control
Staged	10% → 50%	same, per cohort
Full	100%

Note: Any cohort regressing on crash rate or battery drain pauses the rollout automatically Devices only download above a battery threshold or while charging, and delta updates keep the bytes down A/B slots mean a bad update rolls back on its own, the rollout gates are about not needing that

Risks

Risk	Mitigation
SOC vendor BSP quality	evaluate two SOCs early
Battery life misses target	measure from the first prototype
Wake latency	resume, don’t boot
Weak Wi-Fi	local clip buffer, RSSI telemetry
RF certification	pre-certified Wi-Fi module
Fleet regressions	staged OTA, A/B rollback

Note: BSP quality is the biggest schedule risk for camera SOCs, kernel age and ISP blobs vary a lot between vendors Battery life is the product, so it’s measured on every build, not at the end

–

Trade-offs

Choice	Gains	Costs
PIR wakes radar	fewer false alerts	~100 ms
Radar only	catches still people	standby draw
Lights on motion	color night video	battery
Longer clips	more context	battery, upload
Keep-alive Wi-Fi radio	instant wake	BOM, 2nd radio
On-device inference	smarter alerts	NPU power, BOM

Note: Every row is a battery-life knob, which is why the metrics are per configuration The defaults are a product decision, fleet data tells us whether they were the right ones

–

Delivery Plan

Phase	Outcome
EVT, ~2 mo	dev kits, BSP up, power measured
DVT, ~3 mo	custom board, all daemons, OTA
PVT, ~3 mo	certification, ~100 device pilot
Launch	staged rollout, dashboards live

Note: Rough durations, mostly gated by the board spins and certification Telemetry ships in EVT firmware, the pilot fleet is where the targets get validated

Summary

The MCU decides when the SOC wakes
PIR + radar: wake for people only
Yocto, signed A/B, sandboxed daemons
≥ 3 months, < 5 s to alert
Telemetry from day one gates every OTA

Note: If Acme remembers one thing: battery life is designed in, measured on every build, and proven across the fleet

Questions?

Appendix

Component details and types

Component : Broker

MQTT (mosquitto) on a Unix socket

Note: on unix socket for the ability to have granular security No TCP listener at all, so nothing on the network can reach it The socket’s file permissions decide who can connect, and the peer uid (SO_PEERCRED) maps each connection to a component without passwords or certs A small auth plugin applies per-topic ACLs to that identity

–

Topics

Topic	Publisher	Subscribers
`motion/event`	mcuctl	cameractl, powerctl
`power/reading`	powerctl	portalctl, telemetryctl
`config/request`	portalctl	configctl
`config/<area>`	configctl	everyone
`camera/clip`	cameractl	telemetryctl
`metrics/<unit>`	everyone	telemetryctl

Note: config/ is retained, so a restarted component gets the current config the moment it subscribes Only configctl can publish config, only portalctl can request changes, the ACLs enforce the arrows in this table

Component : MCUctl

Owns the MCU link: motion, lighting and the BLE radio

Note: Talks to MCU and sets/gets the motion detection and light settings

–

Motion

Which sensors watch for motion

enum MotionDetectionMode {
    OnlyPIR,
    PIRWakesRadar,
    OnlyRadar,
    Off
}

–

Lighting

When the visible and IR spots come on

enum LightControlMode {
    DuskDawn,
    Motion{cooldown: u32},
    Off
}

Component: Powerctl

Power modes, schedules and PMIC telemetry

Note: Collects power usage data, rundown and fuel gauge from the PMIC

–

Power Modes

Awake on detection, or on a schedule

enum PowerCtlMode {
    Detection{cooldown: u32},
    Timer{configs: Vec<TimerConfig>},
}

struct TimerConfig {
    start_time: u64,
    end_time: u64,
    schedule_days: HashSet<Weekday>
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
enum Weekday { M, T, W, R, F, S, U }

–

Power Readings

One sample from the PMIC

enum PowerSource { Battery, External }

struct PowerReading {
    source: PowerSource,
    timestamp: u64,
    voltage_mv: u32,
    current_ma: i32,              // + charge / - drain
    state_of_charge: Option<u8>,  // fuel gauge, %
    time_to_empty_s: Option<u32>, // rundown estimate
}

Component: Portalctl

First-boot setup, then a status portal

Note: Interactive wireless authentication and cloud configuration

–

Modes

Which portal is up, and how setup arrives

enum SetupTransport {
    Ble,
    SoftAp { ssid: String, psk: String },
}

enum PortalMode {
    Setup { via: SetupTransport },
    Status,
    Off,
}

Note: Setup runs on first boot / factory reset, over BLE through mcuctl from the phone app, soft-AP only as a fallback without the app Status once on the home network, Off once cloud-managed

–

Provisioning

How far setup has got

enum ProvisioningState {
    AwaitingCredentials,
    JoiningNetwork,
    Claiming,
    Provisioned,
    Failed { reason: String },
}

Note: The phone app polls this over BLE, or the soft-AP page does, to show progress and failures

–

Credentials

What setup collects from the user

enum WifiSecurity { Open, Wpa2Psk, Wpa3Sae }

struct WifiCredentials {
    ssid: String,
    security: WifiSecurity,
    passphrase: Option<String>,
}

struct CloudClaim {
    claim_token: String, // QR / app
    display_name: String,
}

Note: claim_token comes from a QR code or the app and is exchanged for device certs

–

Config Requests

Changes sent to configctl via the broker

enum ConfigRequest {
    Network(WifiCredentials),
    Cloud(CloudClaim),
    DisplayName(String),
}

Note: portalctl publishes these over the broker; configctl validates and applies them

–

Status Portal

The status portal’s read-only view

enum BootSlot { A, B }

struct DeviceStatus {
    display_name: String,
    firmware_slot: BootSlot,          // A+B boot
    state_of_charge: Option<u8>,      // powerctl
    motion_mode: MotionDetectionMode, // mcuctl
    cloud_connected: bool,
}

Note: Read-only view assembled from the other components via the broker

Component: Cameractl

Frames to the clip buffer, the cloud and inference

Note: Reads frames from the camera and decides to do further processing depending on config

–

Capture

What starts a recording, and how much

enum CaptureTrigger { Motion, LiveView, Timer }

enum Resolution { P720, P1080, P1440 }

struct ClipPolicy {
    resolution: Resolution,
    max_len_s: u16,
    extend_on_motion: bool,
}

Note: extend_on_motion keeps recording while the radar still sees someone, up to max_len_s

–

Clips

Where frames go

enum Sink { ClipBuffer, Cloud, Inference }

struct Clip {
    id: u64,
    trigger: CaptureTrigger,
    started_at: u64,
    len_ms: u32,
    uploaded: bool,
}

Note: Every clip lands in the local buffer first, then uploads, so a weak or missing network never loses one

Component: Configctl

Source of truth for settings

Note: Source of truth for various configurations set from the portal

Network configuration
Display Name
Cloud Configuration
On-Device Processing Configuration

–

Device Config

One versioned document

struct DeviceConfig {
    version: u32,
    display_name: String,
    network: WifiCredentials,
    cloud: CloudConfig,
    motion: MotionDetectionMode,
    light: LightControlMode,
    power: PowerCtlMode,
    processing: ProcessingConfig,
}

Note: Written atomically to /data/config, then published retained on config/ The cloud can push changes too, they go through the same validation as the portal’s

–

Cloud & Processing

struct CloudConfig {
    endpoint: String,
    device_id: String,
}

enum ProcessingConfig {
    Off,
    Detect {
        targets: Vec<InferenceTarget>,
        threshold: f32,
    },
}

Note: ProcessingConfig is what infctl reads to decide which models to run

–

Applying Changes

Every request gets an answer

enum ConfigResult {
    Applied { version: u32 },
    Rejected { reason: String },
}

Note: Every ConfigRequest gets a ConfigResult back, so the portal can show why a change was refused

–

Component: Infctl

On-device inference on the SOC’s NPU

Note: This one is – because it’s a bonus one that we may not get into Configures on-device inference using NPU hardware available in modern camera SOCs

–

Models

struct Model {
    name: String,
    version: String,
    input: (u16, u16),
}

enum InferenceTarget {
    Person, Vehicle, Package, Animal,
}

Note: Models ship with the OS image so they’re signed and A/B like everything else

–

Detections

struct Detection {
    target: InferenceTarget,
    confidence: f32,
    bbox: (u16, u16, u16, u16),
}

Note: A detection above the threshold upgrades a motion event into a person / vehicle / package alert

How I Used AI

Where	How
Slides	drafted text, tables and notes
Types	sketched the Rust types per component
Review	checked the deck against the brief
Tooling	built and deployed the deck site

Note: Draft, edit to match what you actually did Tooling included the Diogramos embed frame mode and theme sync, the Forgejo CI and the Argo deploy

–

Practices

AI drafts, I decide the architecture
Numbers checked by hand: power math, part specs
Small steps, reviewed and committed as I go
Every slide rendered and checked

Note: Draft, edit to match how you actually work