Vistect

Independent Visual Document Understanding and Creation for Blind Professionals

Inspiration

Vistect began with a seven-hour wait.

Arif Hosen is a Project Coordinator at Songjog Foundation, a nonprofit organization, and a member of our team. He is visually impaired and independently handles complex professional responsibilities, including coordinating projects, preparing communications, organizing information, and producing reports.

During one project, Arif needed to create a press release and modify a project report card.

He could write the content, organize the information, review the wording, and determine what the documents needed to communicate. The barrier appeared when he needed to add and modify visual elements, understand their placement, and verify the final layout.

The document tools available to him did not provide sufficient nonvisual access to those visual operations. Arif had to wait for a sighted coworker to become available.

He waited seven hours.

The task itself did not require seven hours. The delay existed because the final visual layer of the work depended on access to someone else's sight.

Arif was fully capable of producing the communication. The limitation was not his knowledge, judgment, creativity, or professional ability. The limitation was an interface that represented images, diagrams, charts, and layouts in a way he could not independently understand or control.

That experience led us to ask:

Why should a capable professional lose hours of productivity, privacy, and creative control simply because document software makes its visual layer inaccessible?

That question became Vistect.

What Is Vistect?

Vistect is a WebMCP-native visual document workspace where blind and visually impaired people can understand existing visual documents and independently create, modify, inspect, and approve professional multipage documents.

It combines two connected capabilities.

Understand

A user can open a visual document and explore:

Pages and sections
Headings and reading order
Images and their purpose
Charts and their underlying values
Diagrams, nodes, connections, and paths
Tables, icons, captions, and callouts
Spatial relationships between visual elements
Visual hierarchy and page composition
Accessibility and layout problems
Uncertain AI interpretations

Create

A user can create or modify a multipage document containing:

Written content
Uploaded images
Contextual alternative text
Icons
Data charts
AI-assisted diagrams
Captions and source notes
Structured page layouts
Accessible reading order

The user can then inspect the result using the same understanding system.

This creates a continuous workflow:

Understand → Create → Inspect → Validate → Correct → Approve → Export

Vistect does not simply generate a document and ask the user to trust it. It gives the user structured access to what was created, how it was arranged, which decisions the agent made, which findings are uncertain, and what remains to be approved.

The Independence Vistect Creates

Before Vistect

Arif could independently:

Write a press release
Organize its message
Edit report content
Review the text
Determine what the document needed to communicate
Make professional and editorial decisions

However, he could still require sighted assistance to:

Add or replace an image
Understand where visual elements were positioned
Select an appropriate photograph
Modify a diagram
Choose and inspect a chart
Check spacing and alignment
Detect overlapping or truncated content
Verify that the exported document looked correct

With Vistect

A blind or visually impaired professional can:

Understand the semantic and spatial structure of a multipage document
Navigate by page, heading, image, chart, diagram, table, warning, or decision
Inspect every important visual element
Upload and compare image candidates
Select images based on purpose, composition, tone, and relevance
Crop and position images using semantic instructions
Create and modify structured diagrams
Trace paths and detect missing diagram connections
Import data and create accessible charts
Verify that a chart accurately represents its source values
Detect measurable layout and accessibility problems
Review changes made by an agent
Approve or reject individual visual decisions
Export the exact document version that was reviewed

The transformation is not merely that document creation becomes faster.

The transformation is that a professional task previously blocked until a sighted person became available can now be completed with much greater independence.

Built With Lived Experience

Arif is not merely the inspiration for Vistect. He is a member of the team whose workplace experience helped define the product.

His experience changed our understanding of the problem.

At first, it was easy to assume that the main challenge was generating visual elements. However, generation alone does not create independence.

An AI can generate an image, chart, diagram, or document layout. But a blind user must still be able to answer:

What exactly did the AI create?
Why was this image selected?
Does the image support my intended message?
What is uncertain about its description?
How are the objects arranged?
Is every diagram node connected?
Does the chart honestly communicate the data?
Is any text overlapping or cut off?
What did the agent change?
Which visual decisions have I not approved?
Does the exported version match what I reviewed?

Vistect is designed around those questions.

The goal is not to let AI make visual decisions for Arif. The goal is to provide enough structured understanding, operational control, validation, and transparency for Arif to make those decisions himself.

“I was able to write the press release and work on the report card content myself. But when I needed to add and modify visual elements, I had to wait for a sighted coworker to help me. I do not need someone else to create the document for me. I need a tool that lets me understand and control its visual elements myself.”

— Arif Hosen, Project Coordinator at Songjog Foundation and Vistect team member

How Vistect Works

Vistect represents a document as a structured semantic model rather than treating it as a flat visual canvas.

The model contains:

Pages
Sections
Text blocks
Images
Icons
Charts
Diagrams
Tables
Captions
Spatial relationships
Reading order
Sources
Accessibility descriptions
Validation findings
Visual decisions
Approval states
Document versions

This allows Vistect to describe and modify the document through meaningful operations.

Instead of requiring the user to manipulate coordinates or inaccessible drag handles, Vistect supports instructions such as:

“Place the image below the introduction.”
“Make the chart the primary visual on this page.”
“Keep the caption with the image.”
“Put the diagram on a separate page.”
“Align these three icons.”
“Connect the unsuccessful outcome back to training.”
“Show every visual decision I have not approved.”

The application translates these semantic instructions into visual operations.

Visual Document Understanding

When a user imports a multipage document, Vistect creates both a semantic and spatial representation.

A semantic page description might say:

Page 4: Employment Outcomes

1. Section heading
2. Introductory paragraph
3. Horizontal bar chart with four categories
4. Chart caption
5. Source note

A spatial description of the same page might say:

The heading appears at the upper left.

The introductory paragraph occupies the upper third of the page.

The chart is the dominant visual element and fills most of the
center and lower page.

The source note appears directly below the chart in smaller text.

This helps the user understand both the content and its visual presentation.

Vistect also communicates uncertainty rather than presenting every AI-generated interpretation as fact.

For example:

High-confidence observation:
Three adults are seated around a table.

Interpretation:
The setting appears professional and collaborative.

Uncertain observation:
One analysis detected a mobility aid near the left edge.
Another analysis classified the object as office furniture.

Recommendation:
Do not include that detail in the final alternative text without
additional verification.

Image Selection and Placement

Vistect helps users upload, inspect, compare, select, crop, and place images.

Instead of returning a single image description, Vistect evaluates candidates according to:

Relevance to the document's message
Composition
Emotional tone
Professional quality
Representation
Potential stereotypes
Crop flexibility
Text-placement compatibility
Distracting details
Resolution
Source and licensing information
Confidence and uncertainty

For example, if the document is meant to communicate independence, Vistect can explain that one image presents a participant working independently, while another may unintentionally suggest rescue or charity.

The user makes the final selection.

Image placement also uses semantic instructions. A user can ask Vistect to:

Keep the main subject visible
Preserve all detected faces
Leave space for a title
Remove a distracting object near the edge
Create a square crop
Use the image as a full-width header
Place the image after a participant story

The proposed crop and placement remain reviewable before approval.

AI-Assisted Diagram Creation

The user can describe a process in natural language, and their agent can create a structured diagram using Vistect's WebMCP tools.

For example:

Create a participant journey showing enrollment, assessment, training, internship, employment, and follow-up. If a participant needs additional training after assessment or internship, show a return path.

Vistect represents the result as a graph containing nodes, connections, groups, entry points, decision outcomes, and terminal points.

The user can inspect:

The complete node list
Incoming and outgoing connections
Primary and alternative paths
Decision outcomes
Cycles
Unreachable nodes
Missing connections
Visual arrangement
Connector crossings
Label collisions

The graph engine performs deterministic checks rather than relying only on an AI statement that the diagram “looks correct.”

The user can then direct corrections and export the diagram as structured SVG, a visual image, and an accessible long description.

Accessible Chart Creation

Vistect allows users to import a dataset and create a chart without surrendering control over how the data is presented.

The system evaluates:

The type of data
Whether the data represents categories or time
The number of series
Label lengths
The user's intended message
Accessibility
The possibility of misleading interpretation

For example, Vistect can explain that a line chart may incorrectly imply a continuous time trend when the data represents separate outcome categories.

Each chart includes:

A visual chart
An accessible data table
A narrative interpretation
Source information
Chart integrity findings

Vistect checks that:

Visual values match the source data
Labels and units are present
Category labels are not truncated
Legends match the series
Color is not the only method of distinction
The chart narrative does not contradict the data

Visual Decisions and Final Authority

Every important visual choice can be recorded in the Visual Decision Ledger.

This includes:

Image selection
Image crop
Image placement
Icon choice
Chart type
Diagram structure
Diagram layout
Page template
Visual priority
Reading order
Alternative text
Long description

A decision record can show:

What was proposed
Which alternatives were considered
Why an option was recommended
Which tool or agent created it
Whether the user approved it
Whether later changes made the approval stale

Before export, the user can ask:

“Show every visual decision I have not approved.”

Vistect presents those decisions as a structured review queue.

The document can only be finalized after required decisions and blocking validation issues are resolved.

Why We Used WebMCP

WebMCP is central to Vistect because visual document applications contain complex state that an agent should not have to infer from screenshots or inaccessible canvas controls.

Vistect knows:

Which pages and visual objects exist
How they are positioned
Their reading order
The underlying chart data
The structure of diagrams
The source of images
Current validation findings
Which decisions are approved
Which document version is active

The user's browser agent understands:

The user's current instructions
The intended audience
The desired tone
The conversation context
Previously expressed constraints
Which unresolved issue the user wants to address

WebMCP connects these capabilities through narrowly scoped, structured tools.

Examples include:

get_document_overview
describe_page_layout
inspect_image
compare_image_candidates
propose_image_crop
place_image_relative_to
create_diagram
connect_diagram_nodes
trace_diagram_path
validate_diagram_structure
import_chart_data
recommend_chart_types
validate_chart_integrity
run_layout_checks
run_accessibility_checks
find_unapproved_visual_decisions
preview_export_manifest
lock_document_version
finalize_locked_export

These are not generic “make it look good” commands.

Each tool performs a precise operation and returns structured state that the user can inspect.

WebMCP allows the agent to collaborate with the user inside the document without guessing how to operate a visual editor. The agent can propose and execute changes, while Vistect enforces validation, versioning, approval, and finalization rules.

Human and Agent Collaboration

Vistect divides responsibility deliberately.

Vistect provides

The semantic document model
Page and object geometry
Visual element metadata
Diagram graph structure
Chart data
Deterministic validation
Reading order
Version and approval state
Export rendering

The agent provides

Natural-language interaction
User-context interpretation
Tool selection
Alternative generation
Explanations and recommendations
Coordination across document operations

The user provides

Purpose
Message
Professional judgment
Representation choices
Corrections
Approval
Final authority

This distinction is important.

Vistect is not designed to automate the user out of the creative process. It is designed to make the visual process accessible enough that the user can remain in control.

Privacy and Security

Professional documents may contain confidential project data, personal images, financial information, participant stories, or unpublished organizational material.

Vistect follows a local-first approach wherever practical.

Operations such as document state management, chart generation, diagram structure, layout validation, reading-order checks, version comparison, and export preparation can be performed locally.

If remote AI analysis is required for an image or difficult document region, Vistect is designed to disclose:

Which asset will be transmitted
Why remote processing is necessary
Whether text or people are detected
What information will accompany the request
Whether the user wants to continue
What will be retained

Imported documents and images are treated as untrusted content. Their text cannot redefine Vistect's tools or bypass approval requirements.

Consequential actions such as locking or exporting a document are tied to an exact document version. If the document changes after approval, the affected approval becomes stale and the user must review the change again.

Accessibility

Vistect is designed to work with existing assistive technology rather than replace it.

The application uses:

Semantic HTML
Keyboard-operable controls
Heading and landmark navigation
Predictable focus behavior
Screen-reader announcements for agent actions
Accessible validation messages
Non-color status indicators
Structured object navigation
No drag-only interactions
Scalable text and high-contrast support

After an agent action, Vistect can announce:

Agent action completed.

Image moved below the Executive Summary heading.

Page 2 now has one unapproved visual decision.

The user can navigate the document by page, heading, image, chart, diagram, table, warning, or approval state.

What We Built

During the challenge, we built the foundation of a WebMCP-native visual document workspace, including:

A semantic multipage document model
Visual document understanding
Page and object navigation
Structured image inspection
Visual element placement
AI-assisted diagram generation
Diagram node and connection management
Chart creation from structured data
Layout and accessibility validation
Visual decision tracking
WebMCP tools for document inspection and modification
Human approval before final actions
Multipage document export

Our demonstration focuses on the workflow that inspired the project: creating and modifying a professional project document containing visual elements.

Challenges We Faced

Visual information is not only object recognition

Identifying that an image contains a person, laptop, and desk is not enough. A professional needs to understand why the image is appropriate, how it is composed, what message it communicates, and what remains uncertain.

We addressed this by separating:

Measurable facts
AI interpretations
Uncertain observations
Human decisions

Generation is easier than verification

An agent can generate a diagram quickly, but independent authorship requires the user to inspect its structure, find missing routes, and understand its visual arrangement.

We represented diagrams as graphs and added structured validation.

Visual canvases are coordinate-driven

Traditional design applications depend heavily on dragging, resizing, alignment guides, and spatial manipulation.

We introduced semantic relationships such as:

place after
align with
group with
make primary
keep together
move to next page

Agent actions can reduce user control

If the agent makes too many automatic decisions, the blind user becomes a passive recipient.

We created approval states, decision records, visible agent actions, and version-bound finalization to preserve user authority.

Accessibility must survive asynchronous changes

Agent actions can modify the document without direct keyboard interaction. This creates focus-management and announcement challenges.

We designed action summaries and review queues so users can understand changes without unexpected focus movement.

The product scope is large

Visual documents can contain countless formats and object types. We deliberately focused the initial product on structured multipage professional reports rather than attempting to recreate every document or design application.

What We Learned

The most important lesson was that accessibility is not the same as automation.

An AI that completes a visual task for a blind person may save time, but it does not necessarily create independence.

Independence requires:

Understanding
Meaningful alternatives
Transparent uncertainty
Operational control
Error detection
Reversibility
Approval
Final authority

We also learned that blind users do not need every visual property translated into coordinates. They need visual relationships translated into useful concepts that support real decisions.

Most importantly, we learned that the final accessibility question is not:

“Can the AI create the document?”

It is:

“Can the user understand, correct, approve, and confidently send it without asking a sighted person to check it?”

Impact

Vistect can support independence in:

Employment
Education
Research
Nonprofit work
Entrepreneurship
Communications
Project management
Reporting
Policy development
Client services

We plan to measure success through outcomes such as:

The percentage of visual document tasks completed without sighted assistance
The number of visual decisions personally approved by the user
Agent-generated errors independently detected
Diagram and chart problems corrected
Unapproved changes blocked
User confidence in sending the final document
Time previously lost waiting for visual assistance

For Arif's original task, the baseline was a seven-hour wait.

The long-term goal is not only to reduce that wait. It is to remove the dependency that created it.

What's Next

Our next steps include:

Expanding testing with blind and low-vision professionals
Improving multipage PDF understanding
Strengthening image comparison and uncertainty reporting
Supporting additional diagram types
Adding more accessible chart formats
Improving PDF and accessible HTML export
Adding privacy receipts for remote visual analysis
Expanding screen-reader compatibility testing
Supporting reusable organizational brand systems
Exploring DOCX and presentation export
Developing tactile-ready diagram and chart outputs
Publishing reusable WebMCP accessibility and approval patterns

We also want to explore how Vistect can work alongside existing workplace productivity tools without requiring blind users to abandon the systems their organizations already use.

Final Vision

Vistect started because a capable professional had to wait seven hours for someone else to modify and verify the visual layer of his work.

We believe blind and visually impaired people should not be limited to consuming descriptions of visual content or accepting whatever an AI generates.

They should be able to direct visual communication, compare alternatives, understand structure, detect errors, make informed decisions, and retain final authority over what they publish.

Vistect gives blind professionals independent control over the visual documents they understand, create, modify, and share.

Built With

ai
api
axe-core
chatgpt
crypto
css3
cytoscape.js
dagre
data
html5
indexeddb
javascript
next.js
ocr
openai
pdf.js
playwright
react
svg
typescript
vega-lite
visualization
web
webmcp

Vistect: Independent Visual Authorship