Case Study: All Hub Content - Multi-Agent Visual Co-Creation System

This case study describes the entire UX design process for All Hub Content Lab, a high-fidelity prototype (AI Spike) that simulates a visual co-creation environment for multimodal content.

The goal is to orchestrate a multi-agent system that transforms creators' knowledge into reusable assets, ensuring brand consistency, user trust, and drastically reducing production time.

Real results: Generation of text and images via Google Gemini API.
Flows and simulations: Agent orchestration, governance, usage metrics, and audit logs simulated and validated with users.
Strategic objective: Validate the user experience and design hypothesis, not build an MVP.

01 The Problem: The Quantified Context Crisis

Content creators suffer from choice overload and enormous cognitive load. The current process is a chaos of "task switching" between disconnected tools, a friction that not only consumes time but also destroys the coherence of the original idea.

The fundamental problem is not the lack of powerful tools, but the absence of an ecosystem that intelligently manages context.

Time lost per day

Survey (n=25)

2.1

hours

30% of the productive day

Tools in average stack

Interviews (12 creators)

apps

Chronic cognitive fragmentation

Creators who lose consistency

Survey

92%

Frustration and rework

Abandonment of complex flows

Behavioral analysis

68%

Lost ideas and opportunities

02 Deep User Research

My process started with data, not assumptions, to discover the true "Jobs to be Done." Based on an exhaustive analysis of 138 different sources, this research reveals a systemic crisis affecting the productivity and well-being of digital content creators.

The main finding is that tool fragmentation and the lack of integrated ecosystems are the root cause of a series of critical problems. Creators face cognitive overload and analysis paralysis ("choice overload") due to the overwhelming number of disconnected options.

One of the most impactful insights is the devastating effect of "context switching," which can consume up to 80% of productive time, destroying creative flow and project consistency.

This inefficiency translates into frustration, widespread burnout, and a tangible loss in content quality. The opportunity, therefore, lies not in creating more tools, but in developing an intelligent ecosystem that manages context, reduces mental load, and unifies the creative workflow.

In-depth Research on the Frustrations of Content Creators in Fragmented Digital Ecosystems

2.1. Research Methodology and Tools

Digital ethnography: Qualitative analysis of threads in online communities, using Perplexity and Gemini Deep Research to synthesize discussions and trends.
Qualitative interviews (10 sessions): 10 semi-structured interviews were conducted with a heterogeneous group of users. The profiles ranged from experts with high technical knowledge to completely non-technical users, thus ensuring a comprehensive view of the challenges and needs from multiple perspectives.

2.2. The 5 Main Pain Points

The analysis revealed that manual context management was the root cause of most frustrations.

Massive productivity loss due to "context switching"

Creators lose between 20% and 80% of their productive time switching between applications, losing the creative "flow" state. This goes beyond "copy/pasting"; it's a constant fragmentation of thought that requires "reloading" the mental context over and over again.

Cognitive overload and risk of burnout

The need to process information flows from multiple disconnected tools exceeds human cognitive capacity (5-9 items). This leads to extreme mental fatigue, degradation of work quality, and ultimately, widespread burnout, as evidenced by testimonials.

Tool fragmentation and information silos

There is no "single source of truth." Data, ideas, and brand guidelines are scattered across different platforms, creating silos. This forces manual work duplication and decision-making based on incomplete or outdated information.

Loss of creative and brand consistency

Original ideas and brand identity get "diluted" in the process of transferring between tools. The need to redo pieces is not a mistake, but an inevitable consequence of an ecosystem that cannot maintain conceptual consistency from the initial idea to the final post.

Analysis paralysis ("choice overload")

The excess of available tools paralyzes creators. They spend a disproportionate amount of time evaluating and choosing technology instead of creating, which leads to dissatisfaction and, in some cases, project abandonment.

2.3. User Persona: "The Context Juggler"

To bring the data to life, I created an archetype that encapsulates the discovered frustrations.

2.4. The Right Design Question (Applying Occam's Razor)

❌ Incorrect Question (Technical Focus)

"How can we build a system to connect prompt nodes in sequence?"

✅ Correct Question (Design Mission)

"How do we design an ecosystem where a creator's context (their brief, their style) is not something repeatedly entered, but a persistent asset that guides a team of agents to produce multimodal content that is always on-brand and trustworthy?"

03 The Solution: A High-Fidelity Prototype

All Hub Content Lab is a visual co-creation environment designed to be intuitive, secure, and efficient, applying principles of design and governance from its conception.

3.1. Prototype Scope (Real vs. Simulated)

Node Interface

Real (React Flow)

Validate the usability of the visual canvas.

Content Generation

Real (Google Gemini API)

Validate the impact of speed (Doherty Threshold).

Agent Orchestration

Simulated

Validate the clarity of the flow without building the full backend.

Governance and Auditing

Simulated

Test the user's perception of trust and control.

3.2. Storyboard of the Ideal Experience

This is the journey I designed for Sara, applying UX principles at every step to take her from chaos to creative flow.

Creative Chaos

Sara, overwhelmed, navigates between apps for a single idea. It's a clear case of 'Choice Overload' that stifles her creativity.

The Focus

In All Hub Content, her idea is in one place. We apply the Law of Simplicity with a clean interface that eliminates distractions.

Context is King

She drags the brief to the PDF node, and the system integrates it as a single unit, applying the Law of Common Region.

Visual Orchestration

This applies the Law of Uniform Connectedness: by visually connecting the prompt to the image, it's understood that one generates the other.

The Magic Begins

She hits 'Generate.' The system responds in under 0.4s, meeting the Doherty Threshold to ensure a fluid and frictionless interaction.

The Vision Materialized

In seconds, the text and an image appear. This is the positive 'peak' of the experience according to the Peak-End Rule.

Creative Control

She adds a prompt to adjust the image. The Law of Conservation of Complexity is applied by transferring the difficulty to a simple connection.

Creation in a Flow State

With one click, she gets the perfect image. By removing friction, the design has allowed her to achieve a creative Flow State.

04 System Design: Architecture and Governance

To materialize my vision of a co-creation environment that would eliminate friction and chaos, I made a fundamental architectural decision: instead of a single, monolithic AI model, I designed an ecosystem of specialized agents.

I built this architecture on Google Conversational Agents (Dialogflow CX), adopting its generative paradigm to create a team of AI Playbooks. Each agent I designed has a very specific Goal and a set of Instructions in natural language that define its expertise, allowing for precise, high-quality collaboration that a generalist could never achieve.

The orchestration of this digital team is not a simple chaining of prompts. I designed a dynamic workflow managed by a Director Playbook (the Orchestrator), which invokes the specialists' capabilities through a system of Tools.

This design pattern was my solution to ensure modularity, scalability, and, most importantly, to materialize my design mission: to turn technical complexity into a fluid, intuitive, and powerful user experience.

4.1. The Digital Agent Team: Mission Profiles

I designed each agent with a role that mimics the structure of an elite creative team. Below, I present their mission profiles and the technical implementation I defined for each.

🤖

Orchestrator Agent

"I am the project director. I translate your vision into an action plan and ensure the team executes it perfectly."

I built the Orchestrator as the system's brain using a Routine Playbook. I designed it to be the single point of contact with the user, interpreting their intentions and coordinating the rest of the team through Tools.

Generative Model:

Gemini 1.5 Pro

🔎

Context Analyst Agent

"I turn the informational chaos of your briefs into structured, ready-to-use knowledge."

This agent is a specialized Task Playbook. I equipped it with a Data Store Tool to "read" documents (RAG). I prioritized speed for instant data extraction.

Generative Model:

Gemini 1.5 Flash

🧠

Profile & Memory Agent

"I remember who you are and who you create for, so you never have to repeat yourself."

I conceived it as the system's persistent memory, implemented as a Task Playbook. It invokes an OpenAPI Tool to interact with Firestore, retrieving "Brand Kits" and feedback.

Generative Model:

Gemini 1.5 Flash

✍️

Copywriter Agent

"Give me a clear brief and a style guide, and I'll deliver text that connects and inspires."

I designed this agent as a pure, specialist Task Playbook. Since creativity is the core of the product's value, I selected the most advanced model to ensure the highest level of quality and linguistic nuance.

Generative Model:

Gemini 1.5 Pro

🛡️

Brand Guardian Agent

"I am the guarantor of consistency. Nothing gets released if it's not 100% true to your brand."

This is my pillar for trust. It's a pure reasoning Task Playbook that returns a JSON verdict. Precision here is crucial for high-reliability critical analysis.

Generative Model:

Gemini 1.5 Pro

🎨

Visual Creator Agent

"I transform concepts and words into impactful images that capture the essence of your idea."

I configured it as a Task Playbook whose sole mission is to invoke an OpenAPI Tool connected to Vertex AI Imagen. To manage this call, a fast model is the most efficient option.

Generative Model:

Gemini 1.5 Flash

📜

Governance Auditor Agent

"I am the digital notary. I record every step to ensure total transparency and traceability."

This component is not a conversational Playbook but an OpenAPI Tool. I designed it as a silent logging service that the Orchestrator invokes to write to Firestore, building traceability from the architecture itself.

Generative Model:

N/A (Service)

4.2. The Orchestrator in Action: Anatomy of a Creation Flow

An agent architecture is powerful, but its value lies in how they collaborate. I designed their reasoning flow to ensure that every step is logical, efficient, and transparent.

Below, I break down the sequence the Orchestrator Playbook follows when receiving a request to generate a post, demonstrating how the system transforms intent into a validated result.

Orchestrator's reasoning flow diagram

This diagram shows the "brain" of the Orchestrator Agent: its step-by-step reasoning process to transform a user's request into finished, validated content. Instead of a simple linear process, it's an intelligent workflow with a critical decision point. The flow can be summarized in three main phases:

Phase 1: Deep Understanding (Steps 1-2)

What happens: After receiving the user's request, the Orchestrator does not immediately try to create anything. Its first priority is to understand the context. To do this, it invokes two specialist agents: the Analyst Playbook to read and structure the project brief, and the Memory Playbook to retrieve the brand's style guide.

Key Design Decision: The system is designed to know before it acts. By gathering all necessary context at the beginning, it avoids generating incorrect content, drastically reducing the need for rework.

+Phase 2: Intelligent Creation and Rigorous Validation (Steps 3-5)

What happens: With the full context, the Orchestrator commissions the text creation to the Copywriter Playbook. However, and this is crucial, the generated draft is not shown to the user. Instead, it is sent to the Brand Guardian Playbook for validation.

Key Design Decision: This is the system's quality control point. Here, the flow branches: if the content is approved, the creative process continues. If it's rejected, the production flow stops, and the system switches to a collaborative mode, constructively informing the user of the problem.

Phase 3: Final Production and Logging (Steps 6-8)

What happens: Only if the text has been approved does the Orchestrator commission the image creation to the Visual Playbook. Once all creative assets are ready, it invokes the Audit Service to log the successful operation, ensuring traceability. Finally, it presents the complete content to the user.

Key Design Decision: The system is resource-efficient, as it doesn't waste time or cost on generating an image if the text is not valid. Additionally, the final logging reinforces trust and governance.

4.2.1. The Real Implementation: Configuration in Dialogflow CX

After designing and building each specialist Tool and Playbook, the Agent Overview view in Dialogflow CX gives us the final "org chart" of our AI team. The following image is not a conceptual diagram; it's a real screenshot of the implemented system, showing how the Orchestrator Playbook connects to and directs its direct subordinates.

What this graph brilliantly visualizes is the difference between two types of invocations in the Orchestrator's Instructions:

Visible Connections (Internal Delegation)

The arrows we see are the direct calls from one Playbook to another, representing internal creative collaboration.

${PLAYBOOK:Creador_de_Copy}

${PLAYBOOK:Guardian_de_Marca}

${PLAYBOOK:Creador_Visual}

Invisible Connections (External Capabilities)

The agents that do not appear connected are invoked through tools to interact with external systems.

${TOOL:Context_Extractor}

${TOOL:Memory_Retriever}

${TOOL:Audit_Logger}

This separation is not a limitation but a deliberate design decision that reinforces Tesler's Law. The Orchestrator doesn't need to know how a PDF is read, only that it has a Tool to do so. This modular architecture allows me, for example, to change the database from Firestore to BigQuery by only modifying the Memory_Retriever Tool.

4.2.2. The Director's Mind: Orchestrator Playbook Configuration

If the Agent Overview is the team's org chart, the internal configuration of the Orchestrator Playbook is its strategic playbook. This is where my conversation design is translated into executable instructions for the AI.

Using a combination of natural language and references to Tools and Playbooks, I defined the exact reasoning sequence the agent must follow. The following image shows the final, saved configuration of the Orchestrator Playbook's "Basics" tab.

Orchestrator Configuration in Dialogflow CX

Each line in the Instructions field represents a deliberate design decision to create an intelligent and resilient workflow:

Context First (Lines 2-3)

The Orchestrator's first action is not to create, but to understand. The instructions force it to use the ${TOOL:Context_Extractor} and ${TOOL:Memory_Retriever} Tools to gather all necessary information (the brief and the style guide) before starting any creative task. This "know before you act" design is my direct solution to the "context crisis" problem I identified in user research.

Delegation to Specialists (Lines 4-5)

Once it has the context, the Orchestrator does not try to do the work itself. It delegates creative and validation tasks to the specialist Playbooks, using direct invocations ${PLAYBOOK:Creador_de_Copy} and ${PLAYBOOK:Guardian_de_Marca}. This modular architecture ensures that the best AI "mind" is used for the right task, maximizing the quality of the result.

Conditional Logic (Lines 6-10)

The IF... THEN instructions in natural language teach the Orchestrator to handle both success and failure. If the Brand Guardian's validation is "approved," the flow continues to visual creation with ${PLAYBOOK:Creador_Visual}. If it is "rejected," the flow stops, and the agent is instructed to inform the user of the reason, turning a potential error into a collaborative iteration loop.

Integrated Governance (Line 11)

The final instruction forces the Orchestrator to log every successful operation using the ${TOOL:Audit_Logger}. This step is not optional; it is integrated into the core of the workflow, ensuring that traceability and governance are an intrinsic part of the creation process, not an afterthought.

In essence, these 11 lines of instructions are not just a script; they are the codification of my UX design philosophy for AI: a system that is contextual, modular, resilient, and transparent by design.

4.3. The Technical Blueprint: Digital Twin and UX Laws

For the above flow to become a reality, I designed a technical architecture that functions as a "Digital Twin" of a real creative team.

This blueprint consciously applies Tesler's Law (Conservation of Complexity), absorbing all the heavy lifting (communication between agents, API management) in the backend to offer a radically simple user experience.

My design is also based on the Law of Uniform Connectedness. By allowing the user to visually build flows on the canvas, the relationships between nodes become explicit and logical, reducing cognitive load.

When executing a flow, the Orchestrator handles the complexity while the Generative UI shows progress in real-time (Goal-Gradient Effect), keeping the user motivated and in a Flow State.

Component Architecture Diagram

4.4. From Intelligence to Mastery: Validating Specialization with Fine-Tuning

My design doesn't stop at creating intelligent agents; my vision is to cultivate them until they become masters. To demonstrate that this vision was technically feasible and not just a hypothesis, I ran a pilot of advanced specialization through fine-tuning.

The goal was simple yet ambitious: to take a general-purpose model and, through focused training with high-quality data, transform it into a specialist with measurable and superior expertise for a critical task.

Specialization Pilot: The Birth of the "Brand Guru"

I selected the Brand Guardian Agent as the perfect candidate for this pilot. Its task of content validation is subjective, full of nuances, and fundamental to user trust.

My implementation process:

Base Model: I chose gemini-2.0-flash-lite-001, an efficient and fast model, ideal for validating the process.
Dataset Curation: I manually created a high-quality dataset in .jsonl format, composed of 100 training examples and a separate set of 8 examples for validation.
Tuning in Vertex AI Studio: I launched a "supervised tuning" job, providing the datasets to train and evaluate the model.

The Results: Quantitative Evidence of Mastery

The training was completed successfully, and the resulting metrics not only validated my hypothesis but exceeded expectations. Below, I present and analyze the data.

Analysis of Learning Metrics:

Accuracy: The accuracy graph is the clearest proof of success.
The blue line (Training) shows that the model quickly learned the study material, approaching 100% correctness.
More importantly, the pink line (Validation) demonstrates that the model not only memorized but learned to generalize, reaching and maintaining an accuracy of over 80% on completely new data. This confirms that it can reason about brand rules, not just repeat them.
Loss: The loss graph, which measures the level of error, reinforces this conclusion. Both curves (training and validation) drop drastically and then stabilize: the classic shape of a healthy learning curve indicating that the model efficiently converged to an optimal solution without overfitting.

Checkpoint Analysis:

This table allows us to see the model's progress at each training "epoch." The final result is compelling: at the final default checkpoint (step 40), the model achieved a validation accuracy of 84.3% (0.843).

Achieving this level of reliability with such a compact initial dataset validates the effectiveness and efficiency of the fine-tuning approach.

Dataset Analysis:

Finally, the analysis of the input and output token distribution confirms that the dataset I designed was balanced. There were no excessively long or short examples that could bias the training, which contributed to a stable and efficient learning process.

Pilot Conclusion: A Validated Strategy Ready to Scale

This successful pilot is more than just a technical experiment; it is the practical validation of my architectural vision. It demonstrates that my design of a modular agent ecosystem not only works but is poised to evolve.

The Vision for the Future: A Team of Masters

The Copywriter Agent can be tuned with the company's thousands of top-performing posts, emails, and articles to learn to replicate success.
The Context Analyst can be tuned with hundreds of internal briefs and documents to learn to identify the organization's specific nuances and priorities.
The Visual Creator can learn to generate prompts for the image model that align with the brand's historically most successful visual aesthetics.

My work on this prototype has not only solved user pain points but has established a scalable foundation and a clear plan to create an AI system that becomes wiser and more valuable with every piece of data and every interaction.

4.5. Governance and Security Principles Reflected in the Design

To ensure user trust and control, the prototype's design simulates the implementation of key governance principles, inspired by the most robust industry standards.

Transparency: The user can see the agent flow.

The Assistant's panel displays the agent flow, showing each step of the process.

Control: The user can intervene at any time.

The user can intervene, pause, or edit at any point in the flow.

Brand Safety: Consistency guaranteed.

The Guardian Agent ensures all generated content is consistent.

Traceability: Immutable history for auditing.

The Auditor Agent creates an immutable history for every decision.

Continuous Improvement: Feedback-based learning (RLHF).

The system is designed to learn from user feedback (RLHF).

CIS Google Cloud v2.0.0

Robust security to protect data on Google Cloud.

Model Armor

Filters prompts and responses to prevent security risks.

Zero-Training

Intuitive interface, with no prior training needed.

WCAG 2.1

Guaranteed accessibility for users with disabilities.

i18n-Ready

Multilingual support for a global experience.

4.5.1. Evolutionary Security Architecture

My security strategy was designed in two phases:

Prototype

Security is managed with the native Safety and Instruction Filters of Dialogflow CX, ensuring a secure testing environment against harmful content and prompt manipulation.

Production Vision

For a large-scale deployment, the system will be integrated with Model Armor. This adds a critical layer of enterprise security, including Data Loss Prevention (DLP) to protect sensitive information in briefs, malicious URL detection, and centralized governance through security templates, all monitored from Security Command Center.

4.5.2. "AI Act Ready" Design – Compliance with the European AI Act (2025)

Although All Hub Content Lab is not a high-risk system according to Regulation (EU) 2024/1689 (AI Act), I designed the prototype to exceed the minimum transparency and governance requirements that come into force on August 2, 2025.

AI Act Obligation (Chapter I and III)	Implemented Design Decision
Inform that interaction is with AI	Persistent banner and contextual microcopy: “🤖 All Hub Assistant, Your creative co-pilot with Artificial Intelligence.
Mark all generated content	Automatic "✨ AI Generated – All Hub" tag attached to every text and image.
Explain the automated process	Planned: "How it works" modal accessible from any node: a step-by-step visual tour of the orchestration.
Effective human control	"Emergency Stop" button active in each phase; flow stops if the human requests it.
Traceability and auditing	Auditor Agent automatically records in Firestore: original prompt, agents involved, decisions, and final outputs (immutable timestamp).
Preparation for regulatory scaling	Architecture prepared to integrate Model Armor in production (DLP, malicious URL detection, and centralized security policies).

Mark all generated content

Automatic ✨ AI Generated tag with tooltip "Image generated with Artificial Intelligence"

Effective human control

"Edit / Reject" button active in each phase; flow stops if the human requests it

AI Act Ready

This prototype is designed to comply with the European AI Act from August 2025: transparency, user control, and integrated traceability.

05 Prototyping and Validation (RITE)

I applied the RITE (Rapid Iterative Testing and Evaluation) methodology. The prototype was built using Google Firebase Studio with React Flow for the interactive canvas, simulating hosting and authentication, and the Google Gemini API for real content generation. Metaprompting and error correction were aided by Google AI Studio.

5.1. Evolution of the Prompt Node Design

The node design evolved from a simple text area to a structured interface with clear parameters, improving usability and reducing errors.

5.2. Validation Results: The Measured Impact

The evolution of the prompt node from a "container" to a "toolbox" was not just an aesthetic change; it had a direct and measurable impact on how users interacted with the system. To quantify this impact, we conducted a series of A/B tests and timed tasks comparing both versions of the node.

Test Methodology

Participants: The 15 creators were divided into two groups. Group A (7 users) used the V1 node, and Group B (8 users) used the V2.
Assigned Task: "Create three variations of a prompt for a marketing campaign for a new product. Save the best version for future use."
Key Metrics: Task Time, Success Rate, Usability Errors, and Perceived Satisfaction (1-5).

Quantitative Impact: The Difference Between Showing and Empowering

The results showed that the new features of the V2 Node were not just "nice-to-haves," but crucial elements that resolved real frictions and unlocked efficiency.

V1 vs. V2 Node Performance Comparison

Task Time (s)

Success Rate (%)

Errors

Satisfaction (1-5)

This chart illustrates the superiority of the V2 Node. It shows a drastic reduction in task time and errors, along with a massive increase in success rate and user satisfaction, quantitatively validating the design decisions.

Qualitative Impact: "Now I actually understand how to use it"

The qualitative feedback revealed the "why" behind the numbers. Users of the V2 Node were not only faster, but they felt smarter and more in control.

Finding 1: Direct actions eliminate mental friction

The V2 Node's toolbar completely eliminated the need for users to "guess" how to perform common actions.

💬 (V1): "I spent a while looking for how to copy the text. I ended up using Ctrl+C and Ctrl+V..."

💬 (V2): "Ah, perfect. I see the duplicate icon. I click it, and it's done. I didn't have to think."

Finding 2: Reusability transforms the perception of value

The "Save" icon was a turning point. Users understood that their work was not ephemeral.

"When I saw the save icon, it all clicked. I realized I wasn't just writing a prompt for this one time, but investing in my future work."
— Agency Designer.

Validation Conclusion:

The evolution of the prompt node demonstrates that actively listening to users and translating their needs into direct, accessible functionalities not only improves usability but transforms the entire perception of the product: from a passive tool to an indispensable creative partner.

06 Design for Evolution: Continuous Learning

My final deliverable was more than an interface design; it was a strategy for the system's evolution. An AI system that doesn't learn is destined for obsolescence. Therefore, I designed the feedback mechanisms and the learning loop to ensure that All Hub not only solves today's task but evolves towards wisdom with every interaction.

6.1. Designed Feedback Mechanisms

The system learns in two ways, combining passive and active user signals.

Implicit Feedback (High-Confidence Signal)

If the user decides to use the generated content (e.g., clicks "Export"), it is recorded as a validated success. This type of feedback carries a higher weight, as it indicates real satisfaction with the result and reinforces the patterns that led to it.

Explicit Feedback (RLHF)

Each content generation is accompanied by a simple yet powerful feedback interface. Every user vote is a labeled training data point that feeds the system's contextual memory.

This design turns a subjective reaction into structured data that the system can use to fine-tune its agents' behavior.

6.2. Diagram of the Learning Loop and Ecosystem Evolution

This diagram shows how each user interaction not only solves their immediate need but also enriches the system, creating a virtuous cycle of improvement.

This loop ensures that the system becomes smarter and more personalized with each use, creating a powerful network effect and a sustainable competitive advantage.

07 Conclusion and Next Steps

All Hub Content Lab demonstrates that a high-fidelity prototype, combining real and simulated interactions, is a powerful tool for validating the experience of a complex, ethical, and, above all, evolutionary AI system.

Key Learnings

Visual orchestration is the solution to context fragmentation.
AI speed is not a technical metric; it's a pillar of the user experience.
Trust is built with transparency, control, and a clear path to improvement.

Immediate Next Steps

Integrate a real analytics system (e.g., Hotjar) to capture prototype heatmaps.
Design and prototype the granular control flows and the Emergency Stop button.
Implement the UI for feedback mechanisms and connect them to a test database.
Conduct an external accessibility audit (WCAG 2.1) and prepare for a SOC 2 security review.

Frequently Asked Questions about Conversational AI

Answers about my specialization in conversational artificial intelligence experience design.

Let's talk?

Are you looking for a UX designer for Artificial Intelligence who can help broaden perspectives and reduce biases in conversational AI?

Fill out the form below or, if you prefer, write to me directly at info@josegalan.dev and let's see how we can work together.

Case Study: All Hub Content - Multi-Agent Visual Co-Creation System

01

The Problem: The Quantified Context Crisis

Time lost per day

Tools in average stack

Creators who lose consistency

Abandonment of complex flows

02

Deep User Research

In-depth Research on the Frustrations of Content Creators in Fragmented Digital Ecosystems

Executive Summary

Main Issues and Frustrations

Specific Evidence: Spanish-Speaking Community

Conclusions and Opportunity

Research Sources

2.1. Research Methodology and Tools

2.2. The 5 Main Pain Points

Massive productivity loss due to "context switching"

Cognitive overload and risk of burnout

Tool fragmentation and information silos

Loss of creative and brand consistency

Analysis paralysis ("choice overload")

2.3. User Persona: "The Context Juggler"

2.4. The Right Design Question (Applying Occam's Razor)

❌ Incorrect Question (Technical Focus)

✅ Correct Question (Design Mission)

03

The Solution: A High-Fidelity Prototype

3.1. Prototype Scope (Real vs. Simulated)

Node Interface

Content Generation

Agent Orchestration

Governance and Auditing

3.2. Storyboard of the Ideal Experience

Creative Chaos

The Focus

Context is King

Visual Orchestration

The Magic Begins

The Vision Materialized

Creative Control

Creation in a Flow State

04

System Design: Architecture and Governance

4.1. The Digital Agent Team: Mission Profiles

Orchestrator Agent

Generative Model:

Context Analyst Agent

Generative Model:

Profile & Memory Agent

Generative Model:

Copywriter Agent

Generative Model:

Brand Guardian Agent

Generative Model:

Visual Creator Agent

Generative Model:

Governance Auditor Agent

Generative Model:

4.2. The Orchestrator in Action: Anatomy of a Creation Flow

Phase 1: Deep Understanding (Steps 1-2)

+Phase 2: Intelligent Creation and Rigorous Validation (Steps 3-5)

Phase 3: Final Production and Logging (Steps 6-8)

4.2.1. The Real Implementation: Configuration in Dialogflow CX

Visible Connections (Internal Delegation)

Invisible Connections (External Capabilities)

4.2.2. The Director's Mind: Orchestrator Playbook Configuration

Context First (Lines 2-3)

Delegation to Specialists (Lines 4-5)

Conditional Logic (Lines 6-10)

Integrated Governance (Line 11)

4.3. The Technical Blueprint: Digital Twin and UX Laws

Component Architecture Diagram

4.4. From Intelligence to Mastery: Validating Specialization with Fine-Tuning

Specialization Pilot: The Birth of the "Brand Guru"

My implementation process:

The Results: Quantitative Evidence of Mastery

Analysis of Learning Metrics:

Checkpoint Analysis:

Dataset Analysis: