CV to Job Matching With Vector Search: Recommendations in Seconds

The careers section of my site gathers several hundred remote jobs from three hiring platforms. Browsing them is the easy part; finding the right ones is not.

• Keyword filters miss good matches. A "backend engineer" and an "API developer" can be the same job, and a filter on one word hides the other.
• Three sources, three formats. Each platform describes skills, rates and locations differently, so the same filter does not mean the same thing everywhere.
• Candidates give up. Scrolling through hundreds of listings to find five relevant ones is exactly the task people abandon.
• A CV is sensitive personal data. Any feature that reads it has to be private, minimal and easy to delete, or it should not exist.

I treated matching as a meaning problem rather than a keyword problem, and privacy as a design constraint from the first line.

One common model for three platforms

A daily sync pulls the open jobs from micro1, Turing and Mercor and normalises them into a single format, with a summary and French and Arabic translations for each listing. Closed jobs are pruned at the next run.

Every job becomes a vector

Each open job summary is turned into a 768 dimension embedding with Google's Gemini embedding model and stored in a Firestore vector index. When a job closes, its vector is removed, so recommendations only ever point to listings a candidate can still apply to.

Reading the CV

A PDF goes directly to Gemini Flash-Lite, and a Word file is converted to text first. The model extracts a structured profile: skills, seniority and domains. That profile is embedded as a search query and compared with every job vector by cosine similarity, and the ten closest open jobs come back to the candidate.

Privacy by design

• Files land in a private storage bucket with public access blocked.
• Uploads are limited to three per day per connection to prevent abuse.
• The candidate receives a deletion token that removes the file and the extracted profile in one click.
• Every CV is deleted automatically after 180 days.
• Being contacted about a matching job is an optional choice, never a condition.

The feature is live on the careers section of vaiibe.com.

• Recommendations in seconds: in a production test, a real PDF CV returned ten relevant matches in about six seconds.
• Matches by meaning, not by wording: a senior Python profile surfaces senior Python and backend roles, and a search about contract law surfaces legal jobs even when the listing uses different words.
• Results that last: the recommendations stay available on the candidate's device, who can hide them, restart with a new CV or delete everything.
• Trust built in: private storage, a one click deletion and automatic expiry make the feature usable without asking candidates for blind faith.

The same architecture applies to any catalogue that people struggle to search: products, courses, properties or internal documents. You can try it on the careers page.