VeraGen: A Claude-first AI architecture with permission-aware retrieval
VeraGen partnered with Vention to develop AI capabilities for its multi-product B2B SaaS platform. The company’s product portfolio includes the live VeraGen Research Hub and a Claude-powered workspace for company documents and organizational knowledge that is currently in development.
Vention is building a Claude-first generative AI architecture using Anthropic models through Amazon Bedrock, with security and access controls embedded throughout the solution.
Claude also became part of the engineering workflow. Vention developers used Claude Code alongside project-specific Skills and MCP integrations to speed up software delivery while keeping engineers in control.
Client
VeraGen is an AI company developing products for research, product development, knowledge work, and business operations. VeraGen.ai serves as the company’s main platform, where Cortex and its other AI products will be published.
The company has already released VeraGen Research Hub, an invitation-only, multi-tenant platform built by VeraGen’s own team using Claude. The Hub lets teams use their own documents, data, context, and instructions to produce clickable prototypes, research posts, visualizations, videos, courses, assessments, and automated workflows.
VeraGen uses Research Hub to build and manage prototypes for Cortex and its other products. VeraGen has also used the Hub to create and publish Good News Network, a public-facing website that demonstrates how work produced in the Research Hub can be released as a digital product.
One of the future platform's products, VeraGen Cortex, is an AI-powered workspace where users can create and organize documents, ask Claude questions about the documents in their space, generate new documents, and return to earlier chats through a searchable chat history.
VeraGen is also developing AI capabilities that generate documents based on user requests and available context. Its InForm product turns uploaded documents into structured, editable forms.
Project description
Following the release of Research Hub, VeraGen continued to develop its portfolio with a secure, Claude-powered workspace for business documents and organizational knowledge.
VeraGen partnered with Vention to build the new product from the ground up. Vention’s scope includes conversational document workflows, tenant separation, role-based access controls, and the underlying generative AI architecture.
VeraGen creates and manages product prototypes in the Research Hub, where its internal team and Vention engineers can review the same working concepts. The Hub supports this part of the product process, while the application under development has its own architecture and engineering scope.
VeraGen has built Cortex for use in its insurance business, where it serves as a data source generator and storage layer for RAG. The architecture supports document generation and the adoption of new Claude models without coupling the application code to specific model versions.
Our solution
Vention engineers are developing a shared Claude-first generative AI architecture and using it to build AI capabilities across VeraGen’s platform.
The current implementation combines Claude Sonnet 4.6, Claude Haiku 4.5, and Claude Opus 4.6 through Amazon Bedrock with modular .NET and Angular components, permission-aware knowledge access, real-time AI interactions, structured document processing, and Claude Code-enabled engineering workflows.
A shared Claude-first generative AI architecture
Vention created a shared generative AI layer for the application under development. The layer separates product logic from individual AI providers and model versions and can be reused as VeraGen adds related capabilities.
The architecture uses Anthropic Claude through Amazon Bedrock as its primary generative AI path. A dedicated Anthropic adapter handles model interactions, while the application layer remains separate from provider-specific implementation details.
We also moved model configuration out of the application code. Bedrock model identifiers, pricing, modalities, and capability settings are managed via environment-specific configuration, allowing VeraGen to add or redirect Claude models without rewriting product functionality. Conversations retain stable logical model identifiers to preserve existing data when the underlying model configuration changes.
For workflows that require predictable, machine-readable responses, Vention implemented structured output through Claude’s tool use with forced tool choice. JSON schemas can be supplied as forced tools, so generated responses deserialize directly into strongly typed .NET objects instead of relying on free-form text parsing.
AI-powered knowledge and document workflows
Vention is developing an AI-powered workspace where users can organize business documents and work with the information they contain through Claude.
The planned product includes:
- Document creation and organization within dedicated workspaces
- Document hierarchies and ownership
- Employee and user access management
- Multiple conversations with retained chat history
- Claude-powered interaction with documents in a user’s workspace
- Role-based access to documents and AI functionality
The conversational interface will let users to search for and work with the knowledge available in their environment, reducing the need to manually search through documents.
Vention also engineered the chat architecture for a production SaaS environment. Claude responds, token by token, via SignalR, while events are first stored in Redis so that conversations can resume after reloads or connection interruptions. This design avoids reliance on sticky sessions and supports multi-instance deployments.
AI usage is metered at the message level, including input and output tokens, cost, and separate prompt-cache reads and writes. VeraGen can use the data to track and manage AI usage and costs across the platform.
Permission-aware AI architecture
Vention designed the new workspace with security and tenant isolation built into the architecture, so VeraGen can support multiple companies, users, and access levels within the same platform.
Vention implemented separate organizational environments where customers can invite employees and assign permissions to specific resources. The team extended the same controls to the AI layer, filtering information based on each user’s permissions before it becomes context for Claude.
Retrieval-level authorization prevents Claude from accessing content that a user does not already have permission to read.
AI-powered document and form automation
Vention extended the Claude integration to support structured document workflows alongside conversational knowledge retrieval.
For VeraGen’s InForm product, uploaded PDFs are processed with Amazon Textract and transformed by Claude into editable form schemas. Claude produces structured JSON through the tool-use workflow, which turns the output into a functional form based on the source document.
Users can then modify the schemas in conversation by describing changes in natural language. Claude returns an updated, structured version of the form.
The Vention team is also expanding the solution with AI-powered document generation. Users will be able to ask Claude to create documents, such as proposals or agreements, using relevant information from documents they have permission to access.
AI-enabled software delivery with Claude Code
Claude plays a second role in the VeraGen engagement as part of Vention’s software development workflow.
Vention engineers use Claude Code extensively for implementation while remaining responsible for validating AI-generated code, architecture, and software quality. Engineers must understand every generated change well enough to review and validate it before it becomes part of the product.
To give Claude Code a persistent project context, we codified engineering conventions directly within the repository rather than keeping them only in documentation or individual engineers’ knowledge.
The setup includes:
- A repository-level CLAUDE.md defining project structure, commands, development constraints, and conventions
- Five committed project Skills with approximately 1,000 lines of codified engineering practices covering development, code review, design implementation, local diagnostics, and environment setup
- Seven MCP integrations connecting Claude Code with Figma, Jira, GitLab, Kendo UI documentation, Microsoft Learn, and NuGet
Claude Code can use the project-specific engineering rules and external development context directly within the workflow. For example, it can access a Figma design, support implementation of the corresponding component, and validate the result against the same design source. Engineers can work across the connected tools without continually switching between them.
Key stats
backend and frontend unit tests
lines of Claude Code Skills
MCP integrations supporting AI-enabled software delivery
We brought Vention in to scale our AI engineering capacity, and they've delivered well beyond staffing.
Their engineers came up to speed fast, held their own alongside our internal team, and consistently contributed ideas that made the work better. They operate like partners invested in the outcome, not contractors executing a spec. I'd recommend them without hesitation.”
Paul Janisko
Manager of VeraGen.ai
Results
Vention has established the core AI architecture for VeraGen’s new knowledge and document workspace and continues to develop the product ahead of its planned release on VeraGen.ai.
Key results include:
- One shared generative AI pipeline serving multiple VeraGen products
- Claude standardized as the primary generative AI path through Amazon Bedrock
- Configuration-driven model management that reduces application code changes when models are added or redirected
- 685 automated tests, including 632 backend and 53 frontend unit tests, plus Playwright end-to-end coverage
- Approximately 1,000 lines of engineering conventions codified as reusable Claude Code Skills
- AI-enabled engineering workflows on Cortex using existing MCP integrations with Figma, Jira, GitLab, Postgres, and other tools
- Real-time, replay-safe streaming for AI conversations
- Permission-aware AI retrieval that prevents unauthorized content from entering the model context
- In progress for Cortex: Claude-powered import and export of PDF data
VeraGen uses Research Hub to create and manage prototypes for its upcoming products. Vention is developing a separate application and AI architecture behind its new knowledge workspace, which VeraGen plans to release as Cortex through VeraGen.ai.
Tech stack
Category | Technologies |
|---|---|
AI models | Claude Sonnet 4.6, Claude Opus 4.6, Claude Haiku 4.5 |
AI infrastructure | Anthropic Messages API, Amazon Bedrock, Amazon Textract |
AI-enabled development | Claude Code, project Skills, CLAUDE.md, MCP |
Backend | .NET 10, Clean Architecture, Entity Framework Core, PostgreSQL, Redis, SignalR |
Frontend | Angular 21, Nx, Kendo UI |
Cloud and storage | AWS, Amazon S3 |
Testing | xUnit, Vitest, Playwright |
MCP integrations | Figma, Jira, GitHub, GitLab, Kendo UI, Microsoft Learn, NuGet |
FAQs
How does VeraGen use Anthropic Claude?
Claude serves two roles in the engagement. Within VeraGen’s products, Claude Sonnet and Claude Opus are accessed through Amazon Bedrock to power generative AI capabilities, including conversational knowledge interaction, structured document processing, and form generation.
Vention engineers also use Claude Code as part of the software delivery process, supported by repository-level instructions, project-specific Skills, and MCP integrations that give Claude engineering and design context.
How does Vention prevent Claude from accessing information users should not see?
Vention built access controls directly into the AI workflow so Claude only works with information each user is authorized to see. Before content is retrieved for an AI interaction, the application checks the user’s permissions and excludes restricted resources from Claude’s context.
Why does VeraGen use Claude through Amazon Bedrock?
VeraGen uses Bedrock mainly for compliance and operational reasons: as a HIPAA-governed platform, it needs customer data to stay within an AWS environment already covered by existing compliance agreements, rather than adding a new vendor data path. Routing through Bedrock also consolidates billing into the existing AWS account, avoiding the need to stand up and reconcile a separate billing integration for a second vendor.
Vention built a dedicated Anthropic adapter and configuration-driven model catalog around the integration, so model identifiers, capabilities, pricing information, and other settings can be managed outside the core application logic.
VeraGen can also redirect an environment to a newer Claude model through configuration while retaining stable logical model identifiers for existing conversations.
How is Claude Code used in Vention’s software development process?
Vention engineers use Claude Code for implementation, with engineers validating AI-generated changes. The team provides Claude Code with repository-specific context through CLAUDE.md, five committed project Skills, and seven MCP integrations.
The repository-level resources define development conventions, review rules, design implementation practices, diagnostics, and environment setup alongside the software they govern. Engineers remain responsible for architecture, code review, and validating that changes meet the project’s quality requirements.