Salman Faris

Hi, I'm Salman

Community @

About Screenly

  • Secure digital signage, built on Ubuntu Core
  • Born in 2012 on the very first Raspberry Pi
  • Anthias: #1 open source signage project on GitHub
  • On Ubuntu Core since 2016 (then "Snappy")

About me

  • Maker and open source contributor
  • Electronics, edge computing, IoT

Digital signage, in one slide

Screens in public places that show content managed remotely. Each one is a small computer, often a Raspberry Pi, behind a display.

🍔 Menus

Restaurants and cafés

✈️ Departures

Airports and stations

🏥 Wayfinding

Hospitals and campuses

🏭 Dashboards

Factories and offices

☁️ Dashboardschedule content
→
PlayerPi + Ubuntu Core
→
📺 Screen24/7, unattended

Unattended, always on, hard to reach: the hardest kind of edge device

10,000+

Screenly players run Ubuntu Core today.
On board since 2016, one of the earliest adopters.

From pets to cattle

🐶 Pets

Hand-tended, one of a kind. When sick, you nurse them back to health.

Your Pi on the desk.

🐄 Cattle

Identical, replaceable, known state. When sick, you replace them.

A fleet in the field.

Screenly, 2013: Puppet on ~10,000 Pis  ·  10-minute runs: pull the plug, brick a screen
Lesson: build the OS for cattle

Ubuntu Core in one picture

Your appsnap
Edge AI modelsnap
Dockersnap
snapdinstalls, updates, rolls back
Base snap: core24read-only root filesystem
Kernel snappi-kernel
Gadget snapPi boot config, interfaces
Raspberry Pi 5
  1. Flash and bootRaspberry Pi Imager, Ubuntu Core 24
  2. Everything is a signed snapkernel, gadget, base and apps
  3. Every update is all or nothingfails? it rolls back on its own
  4. Apps are confinedhardware only through interfaces

Choosing an edge OS: the honest comparison

Raspberry Pi OSYocto / BuildrootbalenaOSUbuntu Core
What it isGeneral-purpose DebianBuild-your-own LinuxContainer OS (Yocto-based)Immutable, snap-based Ubuntu
Appsapt packagesWhatever you build inDocker-compatible containersSnaps (Docker as a snap)
OS updatesIn place, no rollbackAdd Mender / RAUC / SWUpdateHost OS updates, rollbackTransactional, auto rollback
App isolationNone by defaultDIYContainersStrict confinement + interfaces
Fleet toolsDIY (Ansible, Puppet)DIY, Mender or hawkBitbalenaCloud / openBalenaSnap Store, Landscape
SupportFollows Debian releasesYou maintain it; LTS up to 4 yrsRolling; ESR lines 9 months10 yrs; 15 with Ubuntu Pro
Best forPrototypesCustom hardware, tiny imagesContainer-first teamsFleets that self-heal for years

Screenly shortlisted Snappy, balena and Yocto in 2016.

Field-ready hardware: what breaks first

💾 Storage fails first

Industrial-grade SD, 16 GB+, Class 10+. Screenly Player Max (x86) uses NVMe; CM4 and CM5 have built-in eMMC.

⏳ Plan the lifespan

Even good cards last ~2 to 3 years under 24/7 writes.

🚀 Better storage

Pi 5 + NVMe HAT, or eMMC on a Compute Module.

🔌 Clean power

Official 27 W supply on Pi 5. Brownouts corrupt storage.

🌡️ Heat

Screens live in hot ceilings and kiosks. Use a heatsink or cooler.

🔐 Know the limits

No TPM on the Pi; RTC needs a battery. High security: x86 + TPM.

Hardware access from a confined snap

  • Your app declares the hardware it needs: plugs and slots
  • Interfaces (camera, gpio, i2c, spi) replace root access
  • One command: snap connect impulse-runner:camera
impulse-runner snapplugs: camera
→
snapd + AppArmorallows only what's connected
→
/dev/video0camera interface
→
📷 USB cameraeverything else: denied

On Pi OS your app is root. On Ubuntu Core it asks for exactly what it needs.

Secure by design

Immutable. Confined. Signed.

  • Immutable: read-only root, signed snaps only
  • Confined: every app in its own sandbox
  • Signed model: defines what a device may run
  • Secure boot + FDE: on TPM/UEFI hardware
  • Pi 5 reality check: no TPM on the board, so no disk encryption
  • Need FDE? Screenly's x86 Max player has a TPM

OTA updates that can't brick a device

  • Kernel, OS and apps update independently
  • Refresh windows and holds: you decide when
  • Validation sets pin a known-good set of snaps
  • Delta downloads keep updates small on slow links
1New revisionarrives from the store
→
2Installed alongsideold revision kept
→
3Restart + health checkhooks verify the app
→
✓ Healthy: commit
✗ Failed: roll back

Managing 10,000 devices without SSH

  • Zero-touch: flash, power on, it runs
  • Your own model + Brand Store decide what ships
  • Landscape or a heartbeat snap for inventory and actions
Your modelsigned list of snaps
→
Brand Storestaged rollouts, validation sets
→
🖥️🖥️🖥️🖥️🖥️🖥️
🖥️🖥️🖥️🖥️🖥️🖥️
10,000 devices
⇄
Fleet viewno SSH, no site visits

Provisioning: from box to screen in minutes

1Flash one imagesame signed image, built in CI
→
2Boot + networkEthernet, or Wi-Fi from network.yaml on USB
→
3Code on screen8 chars, no look-alikes, 30 daysk7qx4ept
→
4Enter in dashboardScreens → Connect screen
→
✓It's cattle nowcontent + signed OTA arrive

One image + a claim code + state in the cloud = replace, don't repair

10,000+ screens today
↓ same fleet tooling ↓
the same playbook for physical AI devices
Same problemsunattended, remote, always on
Same answerssigned OTA, rollback, confinement
New valueprivate AI: data never leaves the device

Physical AI: AI meets the real world

A Gartner top 10 tech trend for 2026

  • What: machines that sense, decide and act
  • Why now: vision models + cheap edge hardware
  • Why the edge: low latency, privacy, offline
  • Why Ubuntu Core: when AI controls a machine, a bad update is a safety problem

Example: factory safety, signage + physical AI

📷 Cameraat the zone entrance
→
🧠 Edge AI snapdetects person, hi-vis vest, hard hat
→
📺 Signage screen"Vest required in Zone B" + live safety stats

🦺 PPE check

A missing vest or helmet triggers an alert on the screen at the door, within a second.

🔒 Private by design

Video never leaves the factory; only counts and alerts are logged.

📈 Same fleet

Screens and AI cameras updated, monitored and rolled back the same way.

Also: line KPIs on the floor screen, forklift-zone alerts, defect spotting at a station.

Edge AI, shipped like a product

Built with Edge Impulse.
Runs on Ubuntu Core.

Train with Edge Impulse

  • Collect and label images, train in Edge Impulse Studio
  • Export one Linux .eim file: DSP + model, optimised for the Pi

Run with impulse-runner (my open source snap)

  • One confined snap runs any Edge Impulse model
  • New model ships as a snap update: signed, with rollback
  • Or bundle app + model into your own custom snap
impulse-runnergithub.com/salmanfarisvp/impulse-runner
edge-vision-coregithub.com/salmanfarisvp/edge-vision-core

Edge Impulse: the ML lifecycle in one place

Edge Impulse Studio web UI · API · Python SDK 1 Collect phone or Pi camera 2 Label AI-assisted labelling 3 Design DSP + learning blocks 4 Train + tune EON Tuner (AutoML) 5 Test live on the Pi 6 Deploy .eim to impulse-runner field data → retrain

🧩 Easy to start

Raw data to a trained model in your browser. No ML expertise needed.

📏 Hardware-aware

Shows latency, RAM and flash for your target before you deploy.

🔁 MLOps built in

One project for data, experiments and deploys. Scriptable via API + SDK.

📦 One file out

A Linux .eim holds DSP + model. It ships as a signed snap update.

Live demo: people detection on a Pi

person · vest 0.94 person · no vest ZONE B ENTRANCE ⚠ 1 vest required Illustration · concept
Factory version: the same pipeline, trained on hi-vis vests
📷 SeeUSB camera on a Raspberry Pi 5
🧠 DecideEdge Impulse person detection model in the impulse-runner snap
📺 Actboxes + live people count, full screen like a signage display
🔄 Shipnew model = signed snap update, with rollback

Demo, recorded: impulse-runner on a Raspberry Pi 5

5 steps beyond the breadboard

  1. Package as a snap from week onesnapcraft init, with strict confinement on from the start
  2. Declare hardware with interfaces, never rootplugs: [camera, gpio] then snap connect
  3. Rehearse revert and recovery before you need themsnap revert my-app, and pull the plug mid-update on purpose
  4. Own your model assertion: the image is the productubuntu-image builds the same signed image in CI
  5. Use industrial SD cards; plan for 2 to 3 years of lifeOr NVMe / eMMC when the device writes all day

Thank you.
Questions?

Salman Faris · Community @ Screenly

QR codeSlides & links
salmanfarisvp.com/
talks/ubucon-india-2026
UbuCon India 2026ScreenlyScreenly