Bad AI In A Box (BAIIAB)

Making AI "Touch Grass"
Tommy Falgout - Partner Solution Architect @ Microsoft
Yui Mikuriya - Partner Solution Architect @ Microsoft

Bad AI In A Box (BAIIAB)

Making AI "Touch Grass"
Lawn Shark

Me

Source: me

WHY?!?!

Make AI Intriguing and Silly, Not Scary

  • Inspired by friends "Fortune Witch"
  • Build upon Adafruit project
  • Wanted to create something tangible and fun
  • AI can feel intimidating and complex
  • Physical interaction makes more approachable

"What if interfacing with AI could be as simple as turning a knob?"

Project Evolution

Started as a learning exercise...

  1. Initial Goal: Learn this newfangled ChatGPT thingy
  2. First Iteration: Bad Advice As A Service (BaaaS)
  3. Physical Interface: Added Raspberry Pi + LCD + Dial
  4. Offline Mode: If WiFi / Tethering fails
  5. Observability: Added OpenTelemetry
  6. Cloud Integration: Connected to Elastic's Observability Platform

The Big Moment

Microsoft Build 2025 - Day 3 Keynote

Co-presented with:

  • Mark Russinovich - Azure CTO
  • Scott Hanselman - Developer Advocate Legend

Demonstrated live on stage:

  • 1000+ live audience
  • 10k YouTube views

User Experience Flow

                                                ┌─────────────────────┐
                                                │   WELCOME SCREEN    │ 4x20 LCD screen
                                                │   AI In A Box       │
                                                └─────────────────────┘
                                                        ↓ (turn to browse)
                                                ┌─────────────────────┐
                                                │   > Joke            │
                                                │     Advice          │
                                                │     Recipe          │
                                                │     Insult          │
                                                └─────────────────────┘
                                                        ↓ (push to select)
                                                ┌─────────────────────┐
                                                │   > Dad Joke        │
                                                │     Funny           │
                                                │     Absurd          │
                                                └─────────────────────┘
                                                        ↓ (push to generate)
                                                ┌─────────────────────┐
                                                │   GENERATING...     │
                                                │   Azure OpenAI GPT-4│
                                                └─────────────────────┘
                                                        ↓ (print)
                                                ┌─────────────────────┐
                                                │       [Logo]        │ 
                                                │      [BAIIAB]       │
                                                │                     │ Receipt Paper
                                                │    [AI Response]    │
                                                │    [Disclaimer]     │
                                                │   [bit.ly/baiiab]   │
                                                └─────────────────────┘

Technical Architecture

DEMO TIME!

Bill of Materials (BOM)

Hardware Components

Component Purpose
Raspberry Pi Zero 2 W 4 core, Debian based
20x4 LCD Display Show navigation menus
Rotary Encoder Navigation (turn + push)
Thermal Printer Receipt paper output
Enclosure Custom 3D printed case

Total Cost: ~$150-$200

The Box

  • Designed in OpenSCAD + Tinkercad
  • Inspired by "The Internet"
  • 3D Printed (PLA)
  • Holes for thermal printer, knobs, display

OpenSCAD

  • 3D Modeler for programmers
  • Enables easy parametric design
  • Customize knob

The Interface

Inspired by my Prusa i3 MK3 printer interface

  • Uses a 20x4 character LCD display
  • Re-wrote interface in Python

Rotary Encoder

  1. Turn Clockwise ↻

    • Navigate down through menu options

  2. Turn Counter-Clockwise ↺

    • Navigate up through menu options

  3. Push Button

    • Select current option
    • Generate AI response

Callbacks handled in Python

That's it! No keyboard, no mouse, no touch screen.

Thermal Printer

  • Originaly from Adafruit
    - Python Library
    - Discontinued!
  • Character Set: ASCII
  • Protocol: TTL Serial, 19200 baud

Battery Capacity

  • mAh = how long a device can run
  • Think of mAh as the size of a fuel tank:
    • A 2000 mAh battery can power a 100 mA device for 20 hours
    • Higher mAh = longer runtime

How do you measure this?!

  • USB multimeters are your friend!
  • Project requires 2 USB connections
    • RPi + Thermal Printer

Content Categories

Stored in config file: Topic -> Subtopic -> Prompt

  • Advice - Bad, Silly, Cryptic, Good
  • Fake Facts - Darth Vader, Satya Nadella
  • Cocktail - Tasty, Disgusting
  • Conspiracy - Funny, Crazy, Dark
  • Insult - Monty Python, Shakespeare, French, German, Spanish

All powered by Azure OpenAI (GPT-4)

Architecture Decisions

Why These Choices?

Decision Reason
Raspberry Pi Affordable, powerful, GPIO support, low power (0.17A)
Python Fast prototyping, rich libraries
20x4 LCD Perfect size, cheap, reliable
3D Print Reproducable, custom design
Rotary Encoder Intuitive, tactile feedback
OpenTelemetry Vendor-neutral, future-proof
Azure OpenAI GPT-4 access, company perk

The Numbers

Project Stats

  • Lines of Code: ~2,500
  • Topics: 5 "main" ones
  • Subtopics: 2-5+ each topic
  • Cost per request: ~$0.02 (GPT-4)
  • Response time: 1-3 seconds average
  • Conference demos: 100+ interactions

Adding Observability

OpenTelemetry Integration

Why? To understand:

  • How users interact with the device
  • API performance
  • Most popular requests
  • How much this is costing Microsoft...

Connected to Elastic's Observability Platform

  • Real-time traces
  • Metrics and dashboards

Elastic Observability Integration

Why Elastic?

Unified Platform:

  • Logs, metrics, traces in one place
  • APM for application performance
  • Custom dashboards and alerts

OTLP Support:

  • Native OpenTelemetry ingestion
  • Easy migration

Metrics Tracked

  • API Calls: Count, duration, model, status
  • User Interactions: Category selections, navigation
  • Response Times: API Call duration
  • Error Rates: Failed calls, timeouts
  • Popular Content: Most requested categories

All visualized in Elastic dashboards

DEMO TIME! (Take 2)

Beyond the Novelty

Physical AI interfaces:

  • Bridge digital-physical divide
  • Make AI accessible to non-technical users
  • Create memorable experiences

Observability:

  • Understand user behavior
  • Debug production issues

Open Source:

  • Share learning pains
  • "With enough eyes, all bugs are shallow..."

Challenges Overcome

Hardware:

  • Make it fit → custom cables + box
  • Periphals → Lots of failing in prod
  • Encoder debouncing → added sleep(?)

Software:

  • API timeout handling → retry logic + offline mode
  • RPi + gRPC = PAIN → Migrate to HTTP (no compilation)
  • Menu navigation logic → state machine pattern

Offline Mode

Resilience Built-In

Fallback responses when network fails:

  • Pre-generated responses
  • Stored in JSON

Benefits:

  • Works at conferences (spotty WiFi)
  • Demo reliability

What's Next?

  • User Feedback (Thumbs up/down)
  • Audio - Not good for conferences
  • Local LLM - 1 Core CPU
  • Mobile? IoT? Camera?
  • Print out cost?
  • Fix issues with some Unicode (e.g. "¾ oz Bourbon")

FAQ

Q: Why Raspberry Pi?
A: Affordable, powerful, great GPIO support, huge community

Q: Can I use different AI models?
A: Yes! Works with any OpenAI-compatible API

Q: Can it work offline?
A: Yes! Fallback mode with pre-generated content

Q: Why OpenTelemetry?
A: Vendor-neutral, future-proof, rich ecosystem

Q: Can I add custom categories?
A: Absolutely! Edit conf/menu.json

Key Takeaways

What Makes This Special

  1. Tangible AI - physical beats digital for engagement
  2. Simple Interface - 3 interactions = maximum accessibility
  3. Observable - understand usage
  4. Resilient - works online and offline
  5. Open Source - learn, share, improve
  6. Affordable - ~$150-200 in parts

Resources

Learn More

Code & Documentation:

Technologies:

BAIIAB - Bad AI In A Box

Slides available at: https://lastcoolnameleft.github.io/baiiab/

Slides writen in MARP

Built with ❤️ and penTelemetry

GitHub: lastcoolnameleft/baiiab
Website: lastcoolnameleft.com
Email: tommy lastcoolnameleft com

My day started at their 2AM With Amanda + Tam Partnership with Ignite BAD != evil. Trying to upend expectations

Don't want you to feel bait and switched My Lawn Shark

BAD != evil. Trying to upend expectations

Inspired by a "Bad advice booth" I attended

Might not seem like a big deal, but I have this as my Team profile pic and still get comments

Show the box

Demo simulator + show code

Design inspired by The Internet from IT Crowd

Demo simulator + show code