Migrate expense AI to Vercel Gateway

[alexradulescu]

BLUF

• Replaces direct Gemini/OpenAI integrations with Vercel AI Gateway.
• Uses Qwen 3.5 Flash as primary and GPT-OSS-20B as automatic fallback.
• Keeps the existing structured-output streaming contract and 300-second Vercel function duration.
• Removes provider-specific SDKs and documents the Gateway environment variable.
• Includes a separate Mantine compatibility fix so TypeScript remains green.

What changed

• Updated both expense categorization API functions to use gateway('alibaba/qwen3.5-flash').
• Configured openai/gpt-oss-20b through providerOptions.gateway.models.
• Removed @ai-sdk/google, @ai-sdk/openai, and the standalone openai package.
• Updated TanStack Router/plugin and Zod versions required for the current build.
• Replaced unsupported responsive props on Mantine Menu with responsive Box wrappers.
• Updated .env.example for VITE_INSTANTDB_APP_ID and AI_GATEWAY_API_KEY.

Validation

• bunx tsc --noEmit — passed.
• bun run lint — passed with 3 existing no-shadow warnings and no errors.
• Targeted formatting check — passed.
• bun run build — passed.
• Browser route sweep across 9 routes — all expected pages rendered with zero console errors.
• API handler validation paths return expected 400/405 responses.

Deployment notes

• AI_GATEWAY_API_KEY must be available in the intended Vercel Preview/Production environment.
• Environment changes require a new deployment.
• Local Vite preview validates the SPA only; Vercel Functions require vercel dev or a deployed preview.

Follow-ups identified during review

These are pre-existing upload-path concerns and are intentionally not expanded into this focused provider migration:

• Return extracted statement text in the body rather than X-Extracted-Text.
• Construct uploaded Blobs from the exact Buffer slice rather than its pooled backing ArrayBuffer.
• Reduce the displayed upload maximum below Vercel's 4.5 MB request limit.
• Add protection/rate limiting for publicly callable AI endpoints.
• Migrate deprecated streamObject to streamText + Output.array before adopting AI SDK 7.

[bolt-new-by-stackblitz]

Run & review this pull request in StackBlitz Codeflow.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
buddy-budget	Ready	Preview, Comment	Jul 11, 2026 6:55am