Case study  /  Platform build, own

A multi-model AI platform for the job search

The job search is research, tailoring, and tracking at volume, exactly the shape of work a multi-model AI platform should handle, if it can isolate each user's data and control cost.

Deployed, in use
  • pgvector RAG, per-user
  • Two model providers
  • 8 job-board integrations
Job Concierge case study card

The engagement

Challenge

High volume, per-role tailoring, and data that must stay isolated

A serious job search is high-volume research plus per-role tailoring plus tracking. A platform can run all three, but only if it keeps each user's data cleanly isolated in retrieval and keeps model cost under control at volume.

Action

A real deployed platform, not a demo

A platform deployed on Vercel with cron jobs and dozens of API routes: per-user-isolated pgvector RAG, prompt caching fed by a stable-versus-volatile context split, two providers (Anthropic and OpenAI), resume and interview tooling, and eight job-board integrations, with background jobs that survive the response.

Result

The search becomes a managed pipeline

A working platform that turns a scattered search into a managed pipeline. The standout is the stable-versus-volatile RAG context split feeding the prompt cache, which is what keeps a multi-model platform both accurate and affordable per user.

How it is built

Retrieval, isolated per user

Your materials
resume, history, targets
per-user RAG→
Multi-model engine
per-user RAG, prompt cache
tailored→
Applications and prep
tailored per role

Per-user pgvector isolation is not a feature here. It is the foundation a shared AI platform needs.

  • 01Isolation. Per-user pgvector namespaces, so one user's retrieval can never surface another's data.
  • 02Cache split. A stable-versus-volatile context split feeds Anthropic prompt caching for accuracy at controlled cost.
  • 03Multi-provider. Anthropic and OpenAI behind a routing layer, chosen per task.
  • 04Background jobs. Long-running generation survives past the HTTP response, with monitoring on the job pipeline.
  • 05Integrations. Eight job-board sources feed the pipeline; reliability-hardened against the documented concurrency edges.

What it proves

Running a real AI product means per-user data isolation and cost control are not features, they are the foundation.

See the rest of the work →

Built around my own career data; a public demo would run on seeded, fabricated data.