Personal AI Job Intelligence System
International job hunters face a problem no existing tool handles well: every market plays by different rules. A Tokyo role reports salary monthly. A Berlin role hides visa sponsorship in paragraph four. A remote role says "worldwide" but means "Americas only." Mainstream trackers like Notion or Huntr let you log what you find manually, but they do not help you find, filter, or evaluate anything.
The result is hours of scanning across sites, a pile of mismatched leads, and no clear signal about which jobs are actually worth pursuing.
The most important product decision was designing the Sovereign Filter as a hard gate that runs before the AI does any work. When the pipeline fetches job listings, this filter checks every one against a personal blocklist of countries, cities, companies, and phrases before any scoring begins. Jobs that fail the check are dropped immediately — never stored, never scored.
The obvious alternative was to score everything first and filter the display afterward. That feels simpler, but it has a real cost: AI scoring is the most expensive step in the pipeline (in time and compute), so running it on 200 jobs just to show 20 wastes 80 percent of the budget on listings that were never going to be relevant.
The smarter approach was to load all the blocking rules in a single database call, then apply them in memory to each job before the AI ever gets involved. That one decision keeps the whole pipeline fast and affordable on a free-tier budget.
A user uploads their CV once. The system breaks it into sections, generates vector embeddings using Voyage AI, and stores them in a PostgreSQL vector database. From there, a single button runs the full pipeline: it pulls listings from 26+ job boards (RemoteOK, Arbeitnow, Himalayas, and more), runs the Sovereign Filter, then scores the surviving candidates by comparing each job description against the stored CV using similarity search and a Groq-powered language model that returns a 0–10 match score with a short explanation.
Salaries are normalized across currencies, Japanese monthly packages are converted to annual with a standard bonus structure, and tax estimates are applied so every number is genuinely comparable. Results appear on a dashboard with filter tabs for remote, relocation, salary, and best match.
This is a fully working personal job intelligence system I built and use in my own search. It replaces three to four hours of manual scanning per week with a single button press. The pipeline runs across EUR, USD, and JPY markets simultaneously, handling currency conversion, visa signals, and regional restrictions in one pass.
It proves I can take a messy, real-world problem, define it precisely enough to build against, and make deliberate product trade-offs that keep the system practical rather than over-engineered.