Architecture Review: Vellum

Vellum claims to be Build AI agents using plain English to automate boring tasks. Let’s look under the hood.

🛠️ The Tech Stack

Vellum is not a simple app; it is a comprehensive LLMOps (Large Language Model Operations) platform. It functions as middleware between your application logic and raw model providers.

• Orchestration Engine: Vellum uses a proprietary visual workflow builder (node-based graph) that compiles down to executable chains. This allows for complex logic (loops, conditionals) that standard “chains” often struggle with.
• Model Agnostic Layer: It abstracts the API differences between OpenAI (GPT-4o), Anthropic (Claude 3.5), Google (Gemini), and open-source models via Groq/Together AI. This allows for hot-swapping models in production without code changes.
• RAG Pipeline: Built-in integrations with vector databases (likely abstracting Pinecone/Milvus) and document loaders. It handles the “chunking” and “embedding” logic which is typically a headache for developers to maintain manually.
• Evaluation Suite: This is the “Deep Tech” differentiator. Vellum includes a regression testing framework for prompts (Unit Tests for AI), allowing developers to run thousands of test cases against a prompt version before deploying.

💰 Pricing Model

Vellum operates on a Freemium model, but with a distinct Enterprise focus.

• Free Tier: Designed for individual developers and prototyping. It typically includes limited monthly credits/runs and access to basic models.
• Growth/Pro Plans: This is where Vellum is known to be steep. Historical pricing for their Pro tier has hovered around $500/month, targeting funded startups and enterprise teams rather than indie hackers.
• Enterprise: Custom pricing for VPC deployments, SSO, and unlimited seats.
• Hidden Costs: You still pay for your own model usage (API keys) on top of the platform fee in many configurations.

⚖️ Architect’s Verdict

Deep Tech (Infrastructure).

Vellum is Production Ready. It is not a “wrapper” in the pejorative sense; it is infrastructure. While a “wrapper” creates a thin UI over ChatGPT, Vellum provides the tooling required to build robust wrappers. It solves the critical engineering challenges of the AI era: version control for prompts, quantitative evaluation of model outputs, and reliable orchestration.

For a solo developer, the pricing might be prohibitive. However, for a team building an AI-native product, Vellum replaces the need to build an internal “AI Platform” team. The new “Plain English to Agent” feature is a low-code entry point, but the real power lies in the underlying developer platform.