Relia - built on sanity

This is a submission for the Sanity Challenge, Path One: Ship an Agent That Queries Real Content.

What I Built
Relia is an evidence-backed upgrade risk investigator. It checks a proposed software upgrade against versioned requirements, compatibility rules, breaking changes, migrations, and exceptions stored as structured Sanity content.
Instead of returning a bare “safe” or “blocked,” Relia follows relationships between the proposed target version and the technologies in the developer’s stack. It produces a decision with cited evidence, a traversable proof trace, and an independent red-team pass. When the evidence does not settle a condition, Relia says so.
The initial knowledge scope covers Next.js 15 → 16, React, React DOM, Node.js, and TypeScript. For example, an investigation can identify that
Node.js 20.8 falls below the documented Node.js floor for Next.js 16, while preserving a separate unresolved question when the supplied stack lacks information needed to evaluate a conditional rule.

Demo

Try Relia here - reliaxbt.vercel.app
For a quick walkthrough, investigate Next.js 15 → 16 with Node.js 20.8. The result shows the decision, source-backed evidence, proof trace, and red-team status.

Code
Relia on GitHub - https://github.com/danielamodu/relia

Relia on GitHub

How I Used Sanity
I modeled the knowledge base as structured, linked Sanity documents: Technologies, Versions, Requirements, Compatibility Rules, Breaking Changes, Migrations, Exceptions, and Sources. The initial facts come from official documentation and release material, including the Next.js upgrade guides, Node.js release information, the React 19 upgrade guide, and TypeScript release notes.
Sanity Context is connected to the production dataset in GROQ mode. I used initial_context and schema_explorer to inspect the connected dataset and schema, and groq_query to retrieve structured records. In production, each investigation sends a contract-scoped GROQ query for the proposed target version, technologies named in the stack, related rules and migrations, and their source records.
Relia then evaluates those retrieved relationships, checks its decision in a red-team pass, and returns a machine-readable artifact containing the evidence, provenance, relationships, and proof trace. It does not silently substitute local fixtures when live retrieval fails.
The reviewed 12-case pilot benchmark scored 100% for Relia and 66.7% for keyword/flat retrieval on that same small case set. This is an early, bounded benchmark—not a claim of general compatibility coverage or a live-network performance comparison.

Sanity Project Details

• Project ID: gjy7dyq2
• Dataset: production
• Studio: relia.sanity.studio

The knowledge base is scoped to the technologies and upgrade path described above; it does not claim complete coverage of every framework version or every project-specific migration.