Please enable JavaScript.

Coggle requires JavaScript to display documents.

WEEK 12 - Agentic AI & Org Transformation, AI DISCLOSURE PERCENTAGE…

- - - - Organizes all business data
      - Uses Claude Code to automate routine data engineering tasks and infrastructure debugging
    - - Safely and efficiently handing control of infra debugging and data pipelines to non‑technical staff
    - - Kubernetes debugging using dashboard screenshots to resolve IP exhaustion
      - Translating plain text into fully automated data workflows for the finance team
      - Helping new hires navigate massive codebases and pipeline dependencies
      - Parallel task management and end‑of‑session documentation updates
    - - Resolved infrastructure problems without specialized network expertise
      - Accelerated onboarding for new data analysts
      - Enabled cross‑team self‑service for finance teams with no coding experience
  - - - Agile Development
      - Stage-Gate Model
      - Design Thinking
      - Lean Startup
    - - Faster innovation
      - Lower failure rates
      - Better customer satisfaction
      - Competitive advantage
    - - Rapid prototyping with autonomous iteration
      - Automated coding and testing
      - Bug fixing and feature implementation
      - Easier onboarding in complex codebases
    - - Reduced manual coding work
      - Improved development speed
      - Better code quality through automation
  - - - Batch Inference
      - Edge Inference
      - Real-time Iference
    - - NVIDIA GPUs
      - AI accelerators
        e.g Google TPUs
      - Cloud AI Services
      - Model serving platforms
    - - Cloud
        
        Runs in data centers
        
        Easier scaling
        
        More powerful
        
        Requires internet
      - Edge
        
        Runs on local devices
        
        Faster local responses
        
        Better privacy
        
        Can work offline
    - - code base comprehension
      - fast file navigation during onboarding
      - explaining ML concepts
      - unit test generation
      - kubernetes management
      - cross-language translation
    - - bridges knowledge gaps
      - accelerates architecture
      - shortens learning curve for complex ML & memory systems
    - - accelerated ML learning
        
        research time reduced by 80%
      - Eliminated language barrier
        
        allowing implementation in unfamiliar languages
  - - - Confidentiality
      - Availability
      - Integrity
    - - Vulnerability triaging
      - Architecture and code reviews
      - Infrastructure debugging
      - Security documentation and runbooks
    - - Faster incident resolution
      - Reduced development bottlenecks
      - Shift from reactive to proactive defense
  - - - Feature development with supervised autonomy for small to medium features.
      - Test generation and code review workflow after manual implementation.
      - Codebase comprehension and fast call stack analysis summarizing components.
      - Immediate Kubernetes operations guidance for configuration and deployment.
    - - Checkpoint: Saved state of progress or model weights for easy rollback.
      - RL: Machine learning where models are trained to make decisions via rewards/punishments.
      - Supervised Autonomy: AI working autonomously but with human oversight and correction ability.
      - Rollback: Returning a system or codebase to a previous stable state upon errors.
    - - Enabled an experimental "try and rollback" approach via frequent check-pointing.
      - Documentation acceleration through automatically generated helpful code comments.
    - - Overview: Focusing on efficient sampling and weight transfers, the RL team uses an iterative approach with frequent checkpointing to let AI autonomously build features.
      - Key Takeaway: Adopting an experimental checkpoint-heavy workflow for supervised AI feature development.
    - - Customize Claude.md to prevent repeated tool-calling mistakes (e.g., unnecessary path changes)
      - Use a checkpoint-heavy workflow to easily roll back experiments that don't work out
      - Try one-shot implementation first; if it fails, switch to a collaborative guided approach
  - - - Overview: The Product Design team supports Claude's interfaces.
        Non-developer designers use Claude Code to bridge the gap with engineering, implementing their design vision directly.
      - Key Takeaway: Bridging the traditional handoff gap, enabling direct code translation of design visions
    - - Mockup: Static visual design showing the final look of a software interface
      - State Manage: Controlling UI data states under different user interactions
      - UI Polish: Final fine-tuning of typography, spacing, and animations in UI
      - Design Handoff: The process of designers passing finished visuals to engineers for coding
    - - Direct front-end polish and large state management changes without engineering hand-offs.
      - Rapid interactive prototyping by pasting mockup images directly into Claude Code.
      - Complex copy changes and legal compliance implementation across the entire codebase.
      - Edge case discovery and system architecture mapping during the design phase.
    - - Weeks to hours cycle time: complex updates completed in two 30-min sessions.
      - 2-3x faster execution, with Claude Code becoming a primary design tool alongside Figma.
      - Two distinct user experiences: augmented workflow for devs, and new capabilities for designers.
    - - Get proper setup help from engineers to overcome the initial technical onboarding barrier.
      - Use custom memory files to guide Claude to explain concepts simply and make incremental changes.
      - Leverage image pasting (Cmd+V) for rapid translation from static mockups to interactive prototypes.
  - - - Custom accessibility solution: built a predictive text speech app for family in just one hour
      - Legal workflow automation: built prototype phone trees to connect staff with the right lawyers
      - Team coordination tools automating G Suite updates and tracking legal review status
      - Rapid prototyping for solution validation to show domain experts for immediate feedback
    - - Enabled prototype-driven innovation, sparking AI capability imagination across departments
      - Heightened security/compliance awareness and prioritized building compliance tools
    - - Shadow IT: IT systems built or used without explicit organizational IT approval.
      - Compliance: Ensuring tech integrations align with legal and internal security standards.
      - Predictive Text: Input technology that suggests words based on context.
      - Accessibility: Designing products usable by everyone, including people with disabilities.
    - - Plan extensively in Claude.ai first to generate a step-by-step prompt before coding
      - Work incrementally and visually: use screenshots to show expectations step-by-step
      - Share prototypes despite imperfection to spark innovation across typically siloed departments
    - - Overview: Automated workflows and built accessibility prototypes without traditional dev resources.
      - Key Takeaway: Equipping legal with shadow IT capabilities to build accessibility and automation prototypes.
  - - - 👤 Overview: As a non-technical team of one, Growth Marketing focuses on performance channels, using Claude Code to build agentic workflows that traditionally require significant engineering resources.
      - 🔑 Key Takeaway: Achieves 10x creative output and compresses hours of campaign setup into mere minutes.
    - - 📊 Automated Pipelines: Google Ads generation pipeline parsing CSVs to create targeted ad variations.
      - 🎨 Custom Tools: Figma plugin development for programmatic mass ad creative production.
      - 🧠 Self-Improving Tests: Advanced prompt engineering with memory systems logging hypotheses.
      - 🔗 Direct Analytics: Meta Ads MCP server integration for direct campaign analytics without switching platforms.
    - - ⏱️ Dramatic Time Savings: Ad copy creation reduced from 2 hours to just 15 minutes.
      - 📈 Unprecedented Scale: 10x increase in creative output, enabling experimentation at a massive scale.
    - - 🤖 Target Automation: Identify API-enabled repetitive tasks as prime candidates for automation.
      - 🧩 Specialized Sub-Agents: Break complex workflows into specialized sub-agents (e.g., headline vs. description).
      - 📝 Plan First: Thoroughly brainstorm and prompt plan in Claude.ai before executing in Claude Code.
    - - 🌊 Agentic Flow: Automated multi-step process driven autonomously by AI agents.
      - 🖌️ Figma Plugin: Extensions for the Figma design software used for batch processing.
      - ⚖️ A/B Testing: Comparing two versions of a webpage or ad to see which performs better.
      - 🗄️ Memory System: Mechanism allowing AI to recall past experiments to optimize future output.
  - - - 🛠️ Overview: Build production-quality analytics dashboards without needing full-stack developers.
      - 🔑 Key Takeaway: Empowers data scientists to bypass full-stack barriers autonomously.
    - - 🏗️ Build complex 5,000-line JS/TS React dashboards from scratch.
      - 📈 Create persistent React dashboards (replacing throwaway Jupyter notebooks).
      - 🧹 Handle repetitive refactoring & complex merge conflicts.
      - 🪂 Zero-dependency task delegation in completely unfamiliar codebases.
    - - 🔄 Workflow Shift: Moved from throwaway notebooks to persistent React dashboards.
      - ⏱️ Time Saved: Achieved 2-4x time savings on routine tasks, accelerating evaluation cycles.
    - - 🎰 Slot Machine Approach: Let it run; accept or restart fresh instead of wrestling with corrections.
      - 🛑 Interrupt for Simplicity: Actively guide the model toward simpler approaches.
    - - 📘 TypeScript: A strongly typed programming language building on JavaScript.
      - 📓 Jupyter: Interactive computing environment blending code and rich text.
      - 🪂 Zero-dependency: Delegating tasks directly to AI without prior project knowledge.
      - 🏗️ Refactoring: Improving internal code structure without changing external behavior.
  - - - 🔑 Key Takeaway: Acts as the 'first stop' for unfamiliar code, granting engineers cross-domain debugging confidence.
      - 🌐 Overview: Serves as a guide for system architecture and navigation to bring external knowledge into Claude.
    - - 🗺️ Workflow Planning: First-step planning to identify exact files for bug fixes or new features.
      - ✂️ Zero Copy-Pasting: Eliminates context-switching overhead by asking directly in Claude Code.
      - 🐛 Cross-Domain Debugging: Independent debugging across completely unfamiliar parts of the codebase.
      - 🧪 Model Iteration: Testing through dogfooding the latest research snapshots.
    - - 💪 Increased Confidence: Faster rotation onboarding and confidence in tackling unfamiliar architecture.
      - 😄 Developer Happiness: Enhanced productivity and massively reduced friction in daily workflows.
    - - 🤝 Iterative Partner: Treat it as a collaborator for iteration, not a one-shot perfect solution generator.
      - 🌱 Minimal Start: Start with minimal information and let Claude guide the discovery process.
    - - 🐕 Dogfooding: Using one's own products or services to test and improve them.
      - 🧠 Context Switch: The mental burden and time lost switching between different windows or tasks.
      - 📸 Model Snapshot: A specific version of an AI model saved at a given point in time.
      - 🧭 Codebase Nav: Moving through and understanding the complex connections within large projects.
- - - - Gen AI
        
        Creation & Dialogue
        
        Cons: Lack of Proactive
        
        Relies Human-in-the-loop
      - Provide comprehensive
    - - Agentic AI
        
        Gen AI with planning & action
        
        Goal-oriented
        
        Virtual digital worker
      - Process Planning
    - - Physical AI
        
        Physical Interaction
        
        AI brain with
        
        Multi-model perception
        
        Logic reasoning
        
        RPA software
      - Execute task in reality
- - - - :arrow_right:Disruptive shift
        
        To survive
        
        Define KSAs Models
        
        Knowledge (K)
        
        AI literacy
        
        Algorithmic & cross-disciplinary logic
        
        Data Privacy laws
        
        Skill (S)
        
        Workflow orchestration
        
        API Integration
        
        Hallucination auditing
        
        Attitude (A)
        
        Agile mindset
        
        Collaborate approach
        
        Human‑centric ethics commitment
  - - - Capability requires competency overhaul
- - - - If-else logic trees
  - - - No hard-code rules
        
        Core of ML(Machine Learmig)
      - Data & Stactistic find hidden patterns
  - - - Transdormer architecture by Google
        
        Use 'Attention Mechmism'
        
        Contextual sentence understanding
  - - - Good at processing complex data
  - - - Massive GPUs &Tokens
        
        Build Foundation Cognition
    - - Pre-trained model is trained on high-quality
  - - - Improve 'Alignment' & Guide AI behaviors
        
        Make model safe, polite & non-offensive
  - - - Search for accurate reference before answering
        
        Ensure up-to-date info
        
        Risk of hallucination :arrow_down:
- - - - Uses
        
        Drones + IoT senseors
        
        Precision farming
        
        Yield prediction
        
        Automated irrigation
      - Goal
        
        Digitize physical environments
        
        Improve efficiency
    - - Uses
        
        ERP integration
        
        Supply Chain monitoring
        
        Automated ordering
        
        Vision AI defect detection
      - Benefits
        
        Reduced downtime
        
        Better Production quality
    - - Uses
        
        Customer service AI
        
        Multilingual support
        
        Financial compliance checking
        
        Legal risk analysis
      - Benefits
        
        Faster operations
        
        Better customer experience
        
        Improved compliance
  - - - Strategy
        
        Private Cloud/on-premise AI
        
        Custom open-source fine-tuning
        
        Multi-agent systems
      - Features
        
        AI Centers of Excellence (CoE)
        
        High security
        
        Deep technological advantage
        
        Timeline & Cost: 6-12+ months
        $300,000 to millions
    - - Goal
        
        Fast ROI
        
        Operational Optimization
      - Strategy
        
        Hybrid Cloud systems
        
        Commercial APIs
        
        RAG technology
        
        ERP/CRM automation
      - Timeline & Cost
        
        3-6 months
        
        $15,000-$100,000
    - - Strategy
        
        No-code SaaS tools
        
        Simple automation
      - Uses
        
        Marketing automation
        
        Administrative
      - Benefits
        
        Low cost
        
        Easy setup
        
        Agile operations
      - Timeline
        
        1-4 weeks
  - - - Retrieval-Augmented Generation -> AI reads company databases/documents
    - - Software as a Service -> Cloud-based AI tools
- - - - Low → human-in-the-loop, slow decisions
      - High → fully autonomous agents, rapid execution
    - - Low → weak supervision, few controls
      - High → strong compliance, strict approval flows
  - - - High autonomy + low governance
      - Extreme risk of failure, high systemic risk score
    - - Low autonomy + very high governance
      - Safe but inefficient; underused AI potential
    - - Balanced autonomy and guardrails
      - High innovation, managed risk
- - - - To: Building Agentic Ecosystems
        Reducing Inference Costs
        Making AI commercially scalable
  - - - AGI research
      - Cloud Computing infrastructure
      - APIs and computing power
    - - Azure 2. AWS 3. GCP
    - - Lower API/inference costs
      - Enable Large Scale AI agents
    - - AI becomes cheaper
      - Businesses can automate more tasks
  - - - -Llama Series -Mistral
    - - Edege AI - AI runs locally on devices
      - Small Language Models(SLMs)
        
        Smaller, targeted AI models
        
        Can run offline
    - - Lower latency
      - Lower cloud costs
      - Better privacy
- - - - Real-world experiences with Microsoft, Ford and the US DoD
  - - - Modules 5-8: focuses on deployment strategy (Change Management, Crawl-Walk-Run)
- - - - NOKS
        
        0%
      - 15%
  - - - TINA
        
        0%
      - 10%