Pathfinder: research pathway navigation for a map of science

The situation

The global structure of scientific research is mapped as a set of research communities, tens of thousands of clusters linked by co-citation and topical overlap. A map like that is a powerful object, but as a static visualization it's nearly useless. Nobody stares at a map of thirty thousand communities and picks out a story. You need a way to ask it questions. What links plasma physics to photonics? How did ion traps give rise to a metrology program? Where did quantum sensing really come from? The map has the answers. The interface has to get out of the way.

What we built

Pathfinder is a Nuxt 3 single-page application built on the SciTech Strategies bibliometric map. The user selects an origin research community and a destination. The system traverses the map to find the pathway between them, highlights the selected communities on the visual map, and generates an AI narrative explaining the intellectual chain that connects the two. Each hop on the pathway cites the communities it crosses and the evidence for the transition.

The frontend runs on Nuxt 3 with Pinia stores for node and path state. The server side uses Nitro with Prisma against PostgreSQL for persistence, Upstash Redis for caching, and Clerk for authentication. Narrative generation goes through OpenAI with structured output schemas. We use responses.parse() with Zod text formatters so every generated story conforms to a schema we can audit. Users can save paths and return to them as a research history.

The architecture follows two design principles written into the project's CLAUDE.md. Fail-open for metering and analytics (a broken meter should never block a user action). Fail-closed for auth and validation (throw early, never silently accept a malformed request).

How it's defensible

Every AI-generated narrative is structured output, not free-form prose. The Zod schema constrains what the model can claim and how it cites the underlying communities. Every hop on a pathway ties to a specific research community in the database, not a hallucinated stand-in. The composable for narrative generation follows a strict loading-error pattern. isLoading is true at start and false in finally. Errors are cleared at start and set in catch. console.error runs with a [ComposableName] prefix. No silent drops.

The backend uses validation middleware on every API route. A single wrapper (withAuthValidatedHandler) takes a Zod schema, pulls the Clerk user ID off the event, validates the body, and passes a typed context into the handler. Webhooks and cached endpoints opt out explicitly. Nothing gets into the database without passing through a schema.

What it replaced

Static PDF exports of the bibliometric map. A research analyst who wanted to understand a pathway had to read a paper, make a few hand drawings, and hope they'd covered it.

What a similar engagement looks like

Pathfinder sits inside SciTech Strategies' commercial offering, but the pattern is reusable. A similar engagement for a different map (patent clusters, technology categories, policy concepts) runs 10 to 14 weeks. We ingest the map, wire the pathway finder, build the narrative schema, and deploy the Nuxt app with auth and billing. The deliverable is a running product your users sign in to, not a report.