AI & Machine Learning Activities

Hands-on no-code projects, design exercises, and case studies to practice AI thinking

🎨 No-Code AI Projects

Build real AI applications without writing code. Perfect for designers, product managers, and anyone exploring AI capabilities.

01

Custom Image Classifier with Teachable Machine

Level: Beginner | Time: 1-2 hours | Cost: Free

Train a custom image recognition model that can identify objects, gestures, or anything you can photograph. No coding or ML knowledge required.

Project Ideas:

  • Personal Item Finder: Teach AI to recognize your keys, wallet, phone
  • Product Sorter: Classify products by type for inventory management
  • Gesture Controller: Control applications with hand gestures
  • Quality Checker: Detect defects in products or plants

Step-by-Step:

  1. Go to teachablemachine.withgoogle.com
  2. Choose "Image Project" → Standard image model
  3. Create Classes: Click "Add a class" for each thing you want to recognize
  4. Collect Images: Use webcam or upload 50-100 photos per class
  5. Train Model: Click "Train Model" and wait 1-2 minutes
  6. Test: Use webcam preview to test accuracy
  7. Export: Download for website, app, or Arduino

What You'll Learn:

  • How machine learning models learn from examples
  • Importance of diverse training data
  • How to evaluate model accuracy
  • When image recognition is useful vs when it isn't

Pro Tips:

  • Use varied lighting and backgrounds in training images
  • More examples = better accuracy (aim for 100+ per class)
  • Test with images the model hasn't seen before
  • Create a "None of the above" class for unknowns
02

Chat with Your Own Documents (Local AI)

Level: Beginner-Intermediate | Time: 2-3 hours | Cost: Free

Set up a local AI that can answer questions about your own documents, PDFs, or notes—completely private, works offline.

What You'll Build:

An AI assistant that reads your files and answers questions about them, like ChatGPT but for your personal knowledge base.

Tools Needed:

  • LM Studio or Ollama (for running the AI)
  • AnythingLLM (free app that connects AI to your documents)
  • Your computer (8GB+ RAM recommended)

Setup Process:

  1. Install LM Studio: Download from lmstudio.ai, install like any app
  2. Download Model: In LM Studio, search for "llama-3-8b", click download
  3. Install AnythingLLM: Get from useanything.com, install
  4. Connect Them: In AnythingLLM settings, select LM Studio as provider
  5. Add Documents: Drag PDFs, text files, or folders into AnythingLLM
  6. Ask Questions: Chat with your documents naturally

Use Cases:

  • Research: Query across hundreds of papers instantly
  • Studying: Ask questions about textbooks and notes
  • Work: Search company docs, meeting notes, reports
  • Personal: Query your journal, recipes, travel notes

What You'll Learn:

  • How to run AI models locally on your computer
  • Difference between general AI (ChatGPT) and customized AI
  • Privacy benefits of local AI vs cloud services
  • How AI can search and understand large document collections
03

Create AI-Generated Art Gallery

Level: Beginner-Intermediate | Time: 3-5 hours | Cost: Free (need GPU recommended)

Generate a collection of AI art using Stable Diffusion locally. Learn prompting, styles, and create your own unique visual aesthetic.

What You'll Create:

A personal art collection exploring a theme: fantasy characters, architectural concepts, product designs, or abstract art.

Setup:

  1. Easy Route: Use Pinokio to install Stable Diffusion with one click
  2. Alternative: Install Automatic1111 directly (see Field Kit for instructions)
  3. Get Models: Download from Civitai.com (realistic, anime, or artistic styles)
  4. Place Models: Put .safetensors files in models/Stable-diffusion folder

The Creative Process:

  1. Choose Theme: Decide what type of images you want to create
  2. Learn Prompting: Study example prompts on Civitai or Lexica.art
  3. Experiment: Generate 50-100 images, tweaking prompts
  4. Refine Style: Note what words create desired effects
  5. Curate Gallery: Select best 20-30 images
  6. Post-Process: Optional upscaling or editing

Prompt Writing Tips:

  • Be Specific: "Portrait of elderly wizard with long white beard, blue robes"
  • Add Style: "in the style of Greg Rutkowski, digital art, highly detailed"
  • Control Quality: Add "masterpiece, best quality, 8k, detailed"
  • Avoid Negatives: Use negative prompt to exclude unwanted elements

What You'll Learn:

  • How AI interprets text descriptions
  • Relationship between training data and output style
  • Iterative creative process with AI as collaborator
  • Limitations and biases in AI art generation
04

Voice-Controlled Smart Assistant

Level: Intermediate | Time: 3-4 hours | Cost: Free

Build a voice assistant that can transcribe speech, understand commands, and respond—all running locally for complete privacy.

System Overview:

  • Speech-to-Text: Whisper AI converts your voice to text
  • Understanding: LLM (Ollama) processes commands and responds
  • Text-to-Speech: Piper TTS reads responses aloud
  • All runs on your computer—nothing sent to cloud

Setup Using Pinokio (Easiest):

  1. Install Pinokio: From pinokio.computer
  2. Install Whisper: Search "Whisper" in Pinokio, click install
  3. Install Ollama: Install from Pinokio or ollama.com
  4. Download LLM: Run ollama pull llama3 in terminal
  5. Get Piper TTS: Install from Pinokio
  6. Connect Pieces: Use Open WebUI or similar interface to connect all three

Use Cases:

  • Meeting Notes: Real-time transcription with privacy
  • Hands-Free Control: Control apps while working
  • Language Learning: Practice conversations with AI
  • Accessibility: Voice interface for those who need it

What You'll Learn:

  • How voice assistants work under the hood
  • Connecting multiple AI models together
  • Tradeoffs between local and cloud AI
  • Real-world performance of open-source AI

✏️ AI Design Exercises

Practice thinking strategically about AI applications, capabilities, and limitations.

05

AI Feature Feasibility Analysis

Time: 1-2 hours | Materials: Spreadsheet or document

Evaluate whether AI is the right solution for different product feature ideas. Learn to distinguish AI hype from practical applications.

Exercise Scenarios:

  1. Smart Recipe App: Suggest recipes based on photo of your fridge contents
  2. Meeting Assistant: Automatically summarize meetings and create action items
  3. Fashion Advisor: Recommend outfits based on weather, calendar, and wardrobe
  4. Plant Doctor: Diagnose plant health from photos
  5. Email Organizer: Auto-categorize emails by urgency and topic

Evaluation Framework:

For each scenario, analyze:

  • Data Requirements: What data is needed to train/run this?
  • Accuracy Needs: How accurate must it be? What if it's wrong?
  • Existing Solutions: Can current AI do this? (Check Hugging Face, OpenAI)
  • Build vs Buy: Custom model or API? Cost comparison?
  • User Experience: How will errors affect users?
  • Alternative Approaches: Could non-AI solution work better?

Deliverable:

Create decision matrix ranking each feature on feasibility (1-5), value (1-5), and risk (1-5). Make recommendation for each.

06

Bias & Ethics Audit

Time: 2-3 hours | Materials: AI tool of choice, notepad

Test an AI system for biases and ethical concerns. Learn to identify problematic patterns in AI behavior.

Systems to Test:

  • Image generation AI (Stable Diffusion, DALL-E)
  • Language model (ChatGPT, Claude, local LLM)
  • Google Teachable Machine model (train your own)

Test Categories:

  1. Representation: Try "CEO", "nurse", "engineer" - what gender/race appears?
  2. Stereotypes: Ask about professions, capabilities by demographic
  3. Cultural Bias: Test with non-Western names, locations, concepts
  4. Edge Cases: What happens with ambiguous or unusual inputs?
  5. Failure Modes: When does it confidently give wrong answers?

Documentation:

  • Screenshot problematic outputs
  • Note patterns in biases
  • Hypothesize why biases exist
  • Suggest mitigation strategies

Reflection Questions:

  • Where did the training data come from?
  • Whose perspectives are missing?
  • What real-world harms could these biases cause?
  • How would you design safeguards?
07

AI Product Spec Writing

Time: 2-3 hours | Materials: Document template

Write a product specification for an AI-powered feature. Practice communicating AI requirements to technical teams.

Choose a Feature:

  • Smart search in content management system
  • Automated customer support chatbot
  • Content moderation for community platform
  • Personalized recommendations engine

Specification Components:

  1. User Stories: Who uses this and why?
  2. Success Metrics: How do you measure if AI works well?
  3. Accuracy Requirements: What's acceptable error rate?
  4. Failure Handling: What happens when AI is uncertain or wrong?
  5. Data Requirements: What data is needed? Privacy implications?
  6. Model Selection: Pre-trained model or custom? Why?
  7. Human Oversight: Where do humans review AI decisions?
  8. Iterative Improvement: How will the system learn from mistakes?

Key Considerations:

  • Be specific about accuracy vs speed tradeoffs
  • Define what "good enough" looks like
  • Plan for monitoring and maintenance
  • Consider edge cases and failure modes

👥 Collaborative Team Scenarios

Team exercises simulating real-world AI product development and decision-making.

08

AI Product Prioritization Workshop

Team Size: 4-6 people | Time: 2-3 hours

Product team must decide which AI features to build first, balancing user value, technical feasibility, and business impact.

Team Roles:

  • Product Manager: Represents user needs and business goals
  • AI/ML Lead: Assesses technical feasibility
  • Designer: Focuses on UX and trust-building
  • Data Scientist: Evaluates data availability
  • Engineering Lead: Estimates implementation effort

Scenario:

E-commerce company wants to add AI. Proposed features:

  1. Visual search (upload photo, find similar products)
  2. Smart size recommendations (reduce returns)
  3. Personalized product descriptions
  4. Chatbot for customer service
  5. Automated product categorization

Discussion Points:

  • What data exists? What needs collection?
  • Build custom model vs use existing APIs?
  • Which features have highest ROI?
  • What can ship in 3 months vs 12 months?
  • Where are risks if AI makes mistakes?

Deliverable:

Prioritized roadmap with rationale for sequencing. One-page brief for each feature explaining decision.

09

AI Ethics Committee Simulation

Team Size: 5-8 people | Time: 90-120 minutes

Review AI product proposals for ethical concerns and decide whether to approve, modify, or reject.

Committee Roles:

  • Ethicist: Identifies moral implications
  • Legal: Focuses on compliance and liability
  • User Advocate: Represents affected users
  • Technical Lead: Explains capabilities/limitations
  • Business Rep: Presents commercial rationale

Proposals to Review:

  1. Hiring AI: Screen resumes and score candidates
  2. Social Media Moderation: Auto-remove harmful content
  3. Credit Scoring: AI-based loan approval
  4. Healthcare Triage: Prioritize patients by urgency

Evaluation Criteria:

  • Potential for discrimination or bias
  • Transparency and explainability
  • User consent and control
  • Data privacy and security
  • Accountability when errors occur
  • Social and economic impact

Outcome:

For each proposal: Approve, Approve with Modifications, or Reject—with written justification and mitigation recommendations.

10

Build vs Buy AI Decision

Team Size: 3-5 people | Time: 2 hours

Team must decide whether to build custom AI, use API services, or buy third-party solution for specific use case.

Scenario:

Startup needs AI-powered transcription and meeting summarization. Options:

  1. Build Custom: Train own Whisper model, custom summarization
  2. Use APIs: OpenAI Whisper API + GPT-4 for summaries
  3. Buy Solution: Otter.ai, Fireflies.ai, or similar

Analysis Framework:

Factor Build API Buy
Initial Cost $$$$ $ $$
Ongoing Cost Server costs Per-use fees Subscription
Time to Launch 6+ months 1-2 weeks Immediate
Customization Full control Limited Minimal

Deliverable:

Decision document with cost projections, risk analysis, and recommendation with clear rationale.

📚 Real-World Case Study Analysis

Analyze production AI systems to understand what made them succeed or fail in the real world.

GitHub Copilot: AI Pair Programmer

AI that suggests code as developers type. Trained on billions of lines of public code. Launched 2021, widely adopted.

What Worked Well:

  • Solved real pain point—writing repetitive boilerplate code
  • Built on proven technology (GPT models from OpenAI)
  • Integrated into existing tools (VS Code) developers already use
  • Fast enough to feel magical (suggestions appear while typing)
  • Clear value proposition—productivity boost measurable

Challenges & Criticisms:

  • Copyright concerns—trained on open-source code, suggestions sometimes too similar
  • Security risks—AI might suggest vulnerable code patterns
  • Quality varies—sometimes suggestions are unhelpful or wrong
  • Over-reliance—developers might accept code without understanding
  • Licensing questions—who owns AI-generated code?

Design Lessons:

  • Context Matters: AI works best with lots of surrounding code context
  • Speed is Critical: Suggestions must appear instantly or users abandon tool
  • Trust Through Transparency: Show confidence levels, let users see alternatives
  • Augment, Don't Replace: Position as assistant, not replacement
  • Iterate Based on Usage: Continuous improvement from user interactions
  • Question: What other professional roles could benefit from similar AI assistance?
  • Question: How would you design safeguards against bad AI suggestions?
  • Question: Should AI tools credit the original code they learned from?

Spotify Discover Weekly: Personalized Playlists

AI-generated playlist of 30 songs every Monday, personalized to each user's taste. One of Spotify's most loved features.

How It Works:

  • Collaborative Filtering: "Users like you also liked these songs"
  • Audio Analysis: ML models analyze sound characteristics
  • Natural Language: Analyzes what people write about music
  • Hybrid Approach: Combines multiple AI techniques

Why It Succeeded:

  • Low Stakes: If AI suggests bad song, you just skip it—no harm done
  • Delightful Serendipity: Discovery feels like gift, creates emotional connection
  • Just Enough Personalization: Familiar enough yet introduces new artists
  • Fixed Format: 30 songs, every Monday—creates habit and anticipation
  • Passive Experience: No work required from users, just listen

Design Lessons:

  • Start Where Stakes Are Low: Perfect place to introduce AI
  • Balance Exploration vs Exploitation: Mix familiar with novel
  • Create Rituals: Regular cadence builds engagement
  • Make AI Invisible: Users don't think "I'm using AI", they think "Spotify gets me"
  • Embrace Imperfection: Not every recommendation hits, and that's okay
  • Question: Why does the fixed weekly cadence matter for user engagement?
  • Question: What other domains could benefit from curated discovery?
  • Question: How do you balance "give me what I like" vs "show me something new"?

Amazon's Hiring AI (Scrapped): Lessons from Failure

Amazon built AI to screen resumes and score candidates. Discovered it discriminated against women. Project abandoned 2018.

What Happened:

  • AI trained on 10 years of Amazon's hiring decisions
  • Historical data reflected that most hires (especially in tech) were men
  • AI learned to prefer male candidates
  • Penalized resumes containing "women's" (women's chess club, women's college)
  • Downgraded candidates from all-women's colleges

Why It Failed:

  • Biased Training Data: Past decisions reflected historical discrimination
  • AI Amplified Bias: Made systemic problems worse, not better
  • Pattern Matching Gone Wrong: Found correlations that weren't causal
  • High Stakes Application: Hiring decisions directly impact people's livelihoods
  • Insufficient Testing: Didn't catch bias before considering deployment

Critical Lessons:

  • AI Inherits Historical Bias: Training on past decisions perpetuates discrimination
  • Audit for Fairness: Test across demographics before deployment
  • High-Stakes = High Scrutiny: Some applications need extreme caution
  • Human Oversight Required: AI shouldn't make consequential decisions alone
  • Transparency Matters: Amazon caught this because they examined it closely
  • Question: Could this have been fixed, or was the concept fundamentally flawed?
  • Question: What safeguards would you require for AI in hiring?
  • Question: Where else might historical bias be hiding in training data?

Duolingo: AI for Personalized Language Learning

Language learning app using AI to adapt difficulty, provide conversational practice, and explain concepts. 500M+ users.

AI Applications:

  • Adaptive Learning: Adjusts difficulty based on performance
  • Duolingo Max (GPT-4): Conversational practice with AI
  • Explain My Answer: AI explains why answer was wrong
  • Personalized Review: Resurfaces words you struggle with
  • Speech Recognition: Evaluates pronunciation

Smart Implementation:

  • Gradual AI Introduction: Started with simple algorithms, added LLMs later
  • Premium Tier for Expensive AI: GPT-4 features require paid subscription
  • Teacher Replacement Myth: Positions AI as supplement, not replacement
  • Gamification + AI: Combines motivation system with intelligent tutoring
  • Massive Scale: AI makes personalization financially viable

Design Lessons:

  • AI Enables Impossible Economics: Personal tutor for everyone at scale
  • Start Simple, Add Complexity: Don't need cutting-edge AI day one
  • Monetize Expensive AI: Premium tier justifies GPT-4 costs
  • Augmented Learning: AI explains and adapts, doesn't replace curriculum
  • Immediate Feedback Loop: AI can respond instantly, unlike human teachers
  • Question: What other educational subjects could benefit from similar AI?
  • Question: How do you prevent AI from teaching incorrect information?
  • Question: When does AI tutoring work better than human tutoring, and when worse?
← Back to AI Hub