Jobvori — AI Job Discovery & Application Platform

The Problem

Job seekers were reviewing dozens of postings one by one and filling in the same forms over and over. Fully automated application bots, meanwhile, produced low-quality applications and violated the platforms' terms of service.

The Approach

Designed the flow as 'AI analyses → system prepares → user approves → application is submitted'. Matching isn't left to the LLM: a deterministic weighted blend of six signals with pgvector HNSW cosine similarity sets the score, and the LLM may only adjust the explanation by ±10 at most. Sites whose terms disallow scraping are never called from the server; the extension works inside the user's own session.

Engineering Highlights

• A cost-ladder AI router: cheap model first, retry on invalid JSON, escalate on low confidence, with budgets, a kill switch and per-provider circuit breakers.
• A SKIP LOCKED job queue on Postgres with backoff, dead-letter and idempotency keys, so no second queue system is needed.
• Submission is bound to a hashed snapshot of the approval and a single-use authorisation; CAPTCHA and MFA are never bypassed.
• Argon2id passwords, rotating refresh tokens with reuse detection, TOTP MFA, AES-256-GCM secrets and a hash-chained audit log.