M. Meriç OrduAI Automation Consultant | Computer Engineering Student
Case study

flx: a clean job feed and an AI bidding agent

From noisy Freelancer.com searches to 84 jobs screened in one run

3 min read

This is my own tool, not a client project. Full write-up and code: github.com/Melwis880/freelancer-cli.

The problem

AI and automation jobs on Freelancer.com are buried in noise. A search for n8n or LangChain also returns logo design, translation and data-entry jobs. Finding the few worth bidding on meant running the same searches by hand and re-reading jobs I had already seen.

The next step, an agent that drafts proposals, needed more than a browser tab: clean, structured, de-duplicated data it can trust.

How it works

Part 1: flx, the data layer. A small command-line tool, Python standard library only, that reads job data from the official Freelancer REST API and does nothing else.

  • Read-only by construction. The HTTP client refuses every method but GET, so it cannot bid or message.
  • Polite to the API. Requests at least 1 second apart, backoff and a retry limit on rate limits.
  • One command for the whole routine. flx scan --only-new runs every keyword and skill search, merges and de-duplicates the results, and shows only jobs no earlier scan has shown.
  • Strict JSON for machines. A versioned schema with a fixed set of fields, so an agent can rely on it.

Part 2: the agent on top. A Claude Code agent defined in plain Markdown, with two skills: “screen listings” and “write a proposal”.

  1. It saves the raw scan to a file before reading a single job, so every decision is traceable.
  2. It treats job text as untrusted: hidden Unicode or lines addressed to an AI stop the run.
  3. It gives every job one decision (bid, ask me, or skip) with a one-line reason.
  4. It drafts proposals only for jobs that pass, citing only proof that exists publicly.
  5. It never sends anything. Bidding stays with me.

The result

  • 84 jobs screened in the first agent run, from 11 searches, no failed search.
  • 8 bid, 34 ask me, 42 skip. The 34 “ask me” jobs came back grouped into 9 questions, so I could decide them by group.
  • 5 proposal drafts the same day.
  • Search tuning, measured: replacing loose multi-word keywords with exact skill searches turned a feed that was about half unrelated into one where about 56 of 83 jobs were clearly on target.
  • 123 tests, CI, MIT licensed.

What it does not prove yet: no time comparison (I never screened this volume by hand) and no bid outcomes so far. Both will be added only when they are measured.

Takeaway

Give the agent clean, trustworthy data and a narrow job, and keep the risky action (sending) with a human. That split is what makes the agent safe to run.