Daily Upwork opportunity review: hard rules, a 0 to 100 score and a person who decides
A daily run that pulls the last 24 hours of postings through the official Upwork API, applies hard rejects and a 0 to 100 fit score, and leaves a top-ten shortlist for me to review. It never submits, messages or spends Connects.
Key facts
- 19Saved searches per runQueried sequentially through the official API
- Top 10ShortlistRanked APPLY or REVIEW
Context and problem
Reading every new posting is slow, and most do not fit. I wanted a short list each morning of the few worth my time, built only through the official API and without any automation acting on my account.
Architecture
How to read it. The top row gathers postings; the bottom row turns them into a decision. The profile rules feed the evaluation from below. Everything inside the boundary is read-only against Upwork. The highlighted box is the person: the system never submits a proposal, messages a client or spends Connects.
A scheduled task runs once a day. A Python export refreshes the OAuth2 tokens, queries 19 saved searches through the Upwork GraphQL API for the previous 24 hours, deduplicates by job ID and writes a temporary file. The evaluation step then works in five stages: hard rejects, a fit score from 0 to 100, an APPLY or REVIEW label, a ranking of the top ten, and a draft opening with a proof point for each APPLY job. The result is a fresh Markdown shortlist, and the raw file is deleted after a successful run.
Alongside the daily review there is JobHunter AI, an interactive web app (FastAPI and React) that fetches postings through the same API and scores a single posting against my profile on request. Its screens appear below with posting text redacted.
Decisions and trade-offs
- Rules first, model second. Hard rejects remove clear misfits before any scoring, so the model only judges postings that could fit.
- A fixed, short output. The shortlist is capped at ten and replaced each day; there is no growing backlog to manage.
- Failure keeps the last good report. An authentication or export error stops the run, reports a non-secret error type and leaves the previous shortlist in place.
Governance and cost
The system is read-only against Upwork: no browser scraping, no client contact, no proposals submitted, no Connects spent and no write endpoints. Every application is mine, reviewed and sent personally. Credentials stay in local configuration, and the profile rules (thresholds, pricing and rejects) are not published here.
Outcome
In daily use. It keeps no record of outcomes, so I make no claim about time saved or proposals won.
What comes next
Recording which shortlisted postings I applied to and how they went, so the score can be checked against real results.
Screens
Screens from the working tool. Private details such as amounts, contacts and client posting text are covered or left out. Select a screen to enlarge it.




