How to Build a Capability-First Model Portfolio in TypeScript

Most AI applications select a provider first and design the product around it. A more durable architecture starts with the capability the product needs.
Instead of asking which provider powers your application, ask which capabilities your application consumes: reasoning, coding, vision, extraction, long-context processing or low-cost generation.
Define a capability contract
type Capability =
| "reasoning"
| "coding"
| "vision"
| "extraction"
| "fast-generation";

interface AIRequest {
capability: Capability;
input: string;
priority: "quality" | "speed" | "cost";
}

interface AIResponse {
output: string;
model: string;
latency: number;
estimatedCost: number;
}
The application now depends on a stable capability contract rather than a provider-specific model name.
Create a model portfolio
const portfolio = {
reasoning: ["reasoning-primary", "reasoning-fallback"],
coding: ["coding-primary", "coding-fallback"],
vision: ["vision-primary"],
extraction: ["structured-output-model"],
"fast-generation": ["low-latency-model"]
};
A routing policy can select a model according to quality, latency, availability or cost.
async function execute(request: AIRequest) {
const candidates = portfolio[request.capability];
const model = await selectAvailableModel(candidates, request.priority);

return modelCloud.generate({
model,
input: request.input
});
}
This is an architectural example, not a provider-specific SDK. The important boundary is between product intent and model delivery.
In production, this layer should also handle credentials, usage measurement, cost records, provider availability and fallback behavior.
VectorNode is building this category as an independent model cloud for AI products: infrastructure that turns models from fixed dependencies into a manageable portfolio of capabilities.