About the Project

Inspiration

Most websites show every visitor the same page, even though people arrive with very different intentions. Someone learning about a service needs education. Someone comparing providers needs evidence. Someone ready to act needs confidence, logistics, and a clear next step.

At the same time, a small business's knowledge is rarely in one place. It is scattered across webpages, PDFs, YouTube videos, social media, press coverage, public datasets, and the owner's own expertise.

AI website builders can generate a polished first draft, but they still depend on the owner to know what to ask for. The owner must translate their business into prompts, decide what information matters, verify what the AI invented, and keep the result current as the business changes.

For a small business owner, that shifts the work rather than removing it.

Our goal is different: the agent should discover what the business already knows, it's brand style, make the design and content decisions it can justify, and return to the owner only for facts or approvals that genuinely require human authority.

This inspired All-in-One Living Website Agent: an agentic system that transforms a business's existing knowledge into a different, evidence-grounded experience for every visitor.

We use Fertility Center of Southern California as our pilot, but the underlying architecture is designed to adapt to other industries, visitor journeys, authoritative sources, and proof points.

Why We Built It for Taskmaster

We built this project for the Taskmaster track because it is not simply a chatbot or a one-time website generator.

It is an end-to-end, event-driven workflow that takes responsibility for completing a complex task. The system decides which stages and tools are needed, routes work autonomously, operates beyond the initial conversation, and asks the user only when their knowledge or approval is genuinely required.

Different events trigger different workflows:

• A website URL launches the complete building process.
• Additional documents or media expand the business's knowledge base.
• A campaign link determines the visitor's initial experience and focus.
• A visitor question triggers intent analysis, retrieval, verification, and in-place composition.
• An owner's answer enters the knowledge base and refreshes the available navigation.
• A scheduled event checks the original sources for changes.
• A source update refreshes the knowledge without erasing approved owner decisions.

The owner is not asked to supervise each step. They can start the build, leave the agent working, and return to a composed experience, a transparent activity log, and a focused list of unresolved questions.

What It Does

The experience begins with something any small business owner can do: paste a website URL.

The owner can optionally provide PDFs, YouTube videos, social channels, media coverage, or other documents. They do not need to arrange components, design navigation, select page sections, or write a detailed prompt.

The Builder Agent then:

1. Crawls and classifies the available material.
2. Extracts facts, services, people, testimonials, outcomes, and logistics.
3. Measures the business's existing brand identity.
4. Connects relevant external proof sources when appropriate.
5. Composes experiences for Explore, Compare, and Decide visitor stages.
6. Reviews the result across multiple screen sizes.
7. Returns anything it cannot verify as a specific question for the owner.

In our fertility pilot, the agent read 29 documents and completed the initial build in approximately 412 seconds.

The finished product is not three separate websites. It is one evidence base and one living system whose order, depth, and emphasis change according to visitor intent.

Someone "exploring" receives education first. Someone "comparing" sees differentiators, detailed options, and proof earlier. Someone "ready to decide" receives outcomes, logistics, and the next action without restarting their search.

Designed for Small Business Owners

Most AI website tools begin with a prompt, a template, or a blank canvas. They help the user produce a page, but the user still has to direct the process: choose sections, organize content, select components, and decide what every visitor should see.

Our system begins with the business itself.

It reads the owner's existing website, documents, videos, social content, media coverage, and approved answers. It then organizes that knowledge around visitor intent, composes the appropriate experience, verifies what it can support, and identifies what is missing.

The difference is not simply that our agent designs the interface. It takes responsibility for the reasoning behind it:

• What should this visitor see first?
• Which evidence is relevant to their current intent?
• What information is missing from the business's sources?
• Which next step is supported by material the business actually has?
• When should the system answer, and when should it admit that it does not know?

The owner is not asked to become a better prompt writer or a part-time web designer. They contribute the business knowledge and judgment only they possess; the agent handles discovery, organization, composition, verification, and adaptation.

How We Built It

The system has three main layers.

Builder Agent

The Builder Agent runs as an autonomous workflow. It crawls webpages and linked documents, processes video and social sources, classifies content, extracts structured facts, measures brand identity, builds the navigation graph, composes stage-based scenes, and reviews the rendered result.

Gemini is used where interpretation and judgment are valuable, including document classification, video understanding, scene planning, design planning, and navigation evaluation.

Deterministic code handles tasks where consistency and safety matter more than creativity, including fact extraction, coverage checks, runtime ordering, evidence rules, and application of owner-approved decisions.

Composer Agent

The resident Composer Agent handles visitor interactions.

It classifies intent across three dimensions:

• Stage: Explore, Compare, or Decide
• Topic: What the visitor is trying to understand
• Persona: Who is reading and what they may prioritize

The agent retrieves supporting evidence and streams a grounded response into the current experience.

Every number and claim must remain traceable to a source. If the knowledge base cannot support an answer, the system returns an honest gap rather than generating unsupported content.

Review Console

Autonomy still requires accountable human judgment.

The console lets the owner confirm, edit, or reject extracted copy, outcomes, credentials, testimonials, images, entry strategy, and visual direction.

These decisions are stored as persistent overlays instead of being written back into the generated source material. Future rebuilds can refresh the underlying evidence without erasing what the owner approved.

The console is intentionally not a conventional page builder. It presents focused business decisions: confirm this claim, identify this person, approve this entry rule, answer this missing question, or select a visual direction.

The owner contributes judgment. The agent handles design and assembly.

Navigation Innovation

Navigation chips are created only after the page scenes have been composed.

For each scene, the agent proposes destinations based on topics the scene raises but does not fully answer. A chip survives only when it:

1. Retrieves real supporting material.
2. Adds information beyond what the visitor can already see.
3. Represents a natural next step from the current scene.
4. Does not duplicate another available destination.

The surviving chip is pinned beneath the scene it was evaluated against. When clicked, its verified query composes a cited response directly where the visitor is already reading.

This makes navigation part of the agent's reasoning rather than a manually designed menu.

The system also walks this verified topic graph after each interaction. It considers what the visitor has already seen, which topic they are following, their current stage, and which content types are still missing from the page. Already visited destinations are not repeatedly offered.

Entry Strategy

A visitor often communicates intent before typing anything.

Campaign tags and referring links can tell the agent whether someone is researching a service, comparing costs, looking for evidence, or ready to contact the business.

These signals guide:

• Which visitor stage should lead
• Which scene should receive priority
• Which topic should be emphasized
• How the existing composition should be reordered

The process is deterministic when an approved rule exists, so the same campaign link produces the same initial experience.

When no rule matches, the agent can infer a likely experience for that visit and propose the missing rule to the owner. The inference does not become permanent until the owner approves it.

Evidence and Industry Adaptation

In our fertility pilot, the business's website did not publish outcome figures. The agent connected the clinic to the CDC's public ART registry and used those outcomes as an external proof point.

This is one industry-specific implementation of a reusable pattern.

For another industry, the system could connect to different authoritative datasets, certifications, ratings, regulations, or business-specific evidence. The contract changes, but the underlying principle remains consistent:

• Verify the entity match.
• Preserve the source and context.
• Never display a number without its meaning.
• Never present public evidence as a guarantee.
• Give up rather than guess when a match is ambiguous.

Technology

We built the project using:

• Gemini on Vertex AI for multimodal understanding and agent reasoning
• Google ADK for the resident answer agent
• Vertex AI embeddings for evidence retrieval
• Cloud Run for the web and Composer services
• Cloud Run Jobs for building and refreshing sites
• Cloud Storage for site bundles and build artifacts
• Firestore for sessions, signals, and workflow state
• Cloud Scheduler for recurring source monitoring
• Next.js for the living website and review console
• FastAPI for the Composer service
• Playwright for multi-viewport visual review
• Firebase Hosting for the embeddable experience
• Public industry data sources, including the CDC ART registry for our fertility pilot

Challenges We Faced

Making Honesty Structural

Prompting a model to "be accurate" was not enough.

We introduced coverage gates, citation requirements, source tracking, and deterministic rules that prevent unsupported numbers or claims from reaching the page.

The system does not merely prefer honesty. Unsupported content has no valid path to the visitor experience.

Balancing Autonomy and Control

We wanted the system to be easy for a non-designer without removing the owner's authority.

Our solution was to automate discovery, design, and assembly while exposing focused human gates for facts, images, entry strategy, and visual direction.

Building Useful Navigation

Early navigation suggestions sometimes repeated information already visible or led to weak destinations.

We redesigned chips as verified promises. Every candidate must retrieve supporting evidence, add new ground, and represent a natural next step. The system records why each candidate was accepted, rejected, or renamed.

Keeping Personalization Stable

A personalized experience should not become random.

We freeze the composed scenes and use deterministic rules to reorder them. Model inference is used only when an entry rule does not exist, and its suggestion must be approved before becoming permanent.

What We Learned

The strongest agentic products combine model intelligence with deterministic structure.

We also learned that making AI accessible does not mean exposing more controls. For a small business owner, the best interface may provide fewer, clearer decisions while the agent handles the technical complexity.

Collaboration also should not require the user to supervise every step. The agent should complete as much work as it can independently, then return with precise questions only where human knowledge or authority is genuinely required.

Finally, good design should be encoded into the system rather than regenerated from scratch for every request. The agent adapts a trusted design grammar to the business instead of improvising an entire interface each time.

What's Next

Our pilot currently includes three visual systems, but we plan to encode additional professional design grammars that the agent can reason with.

We also plan to add:

• More industry contracts and visitor journeys
• Additional authoritative data connectors
• Authenticated owner consoles
• Approved social-platform integrations
• More brand-aware UI and motion options
• Recurring monitoring across connected sources
• Selective refreshes that update only affected content and navigation

Our goal is not to give every small business owner another website builder. It is to give them an agent that understands their business, completes the marketing work, and asks for their judgment only when it matters.

Your knowledge in. A living site out. A different all-in-one page for every visitor.

Built With

• cloud-run-and-cloud-run-jobs
• cloud-scheduler
• cloud-storage
• fastapi
• firebase
• firestore
• gemini-on-vertex-ai
• google-adk
• next.js
• playwright
• vertex-ai-embeddings