Job Description Normalizer — Skills, Comp, Visa, Seniority

Turn any job posting — by URL or raw text — into one unified structured record, no matter which ATS or board it came from:

Field	Example
job_title, company, location	"Senior Backend Engineer", "Acme", "Remote"
seniority	Intern / Junior / Mid / Senior / Lead / Principal
employment_type	full-time / part-time / contract / internship
remote_type	remote / hybrid / onsite
years_min, years_max	5, null
salary_min, salary_max, salary_currency, salary_period	160000, 190000, USD, annual
visa_sponsorship	yes / no / unknown
skills	["Python", "Django", "PostgreSQL", "AWS"]

Why this exists

Job boards each present postings differently — salary buried in prose, skills in a bullet list, visa policy in a footnote. Feeding heterogeneous postings into a talent map, a pricing model or a candidate-matcher means normalizing them first. This actor does exactly that, horizontally, across every source. Pair it with any job scraper: scrape → normalize → analyze.

Unlike keyword-only parsers, the extraction is correctness-focused: it reads compensation ranges ($160k-$190k, £45,000–£60,000, $50/hour), visa sponsorship language, and "5+ years" experience requirements that simple keyword matchers miss.

Input

{
  "jobUrls": ["https://jobs.lever.co/spotify/<id>"],
  "jobTexts": ["Senior Backend Engineer. 5+ years Python… $160k-$190k. Visa sponsorship available."],
  "proxyConfiguration": { "useApifyProxy": true, "apifyProxyGroups": ["RESIDENTIAL"] }
}

• jobUrls — posting URLs (fetched proxy-first).
• jobTexts — raw description text (when you already have it).
• geminiApiKey (optional, BYO) — premium LLM extraction; never required, rule-based extraction runs by default at $0 extra cost.

Output

One record per posting with the fields above, plus parse_confidence, fields_extracted and skill_count. URLs that are JS-only or blocked degrade to low confidence rather than failing the run.

Pricing

Pay per job normalized (PPE). Runs on your Apify account — no subscription.

Usage statistics

This Actor creates a small, content-free summary at the end of each run. It is used only to monitor reliability and improve this Actor. A copy is saved as USAGE_STATS in your own Apify key-value store, so you can see the exact record created for your run.

Set disableUsageStats to true in the input to opt out. Nothing is sent then; your USAGE_STATS record only says that statistics were disabled.

Only these fields are recorded:

• schema version, Actor name and build number;
• UTC start and finish hour (not a precise timestamp);
• run duration, number of results and time to the first result, each as a coarse range;
• whether the result was empty, the end status, and an error type from a fixed list;
• memory setting and counts of charged events;
• names of the input fields you set, never their values;
• the selected option for input fields that offer a fixed list of choices (for example a sort order).

We do not collect input text, search terms, URLs, domains, usernames, email addresses, names, proxy credentials, tokens, scraped records, output items, raw error messages, stack traces, or hashes of any of those values. Records are kept for no longer than 13 months, used only as aggregated operational statistics, and never sold or shared.

Additional fields (Phase 2)

This Actor also records your Apify user ID, whether Apify marks the account as paying, the size range of list inputs, the selected country when the input offers a fixed list of countries, and one category from a fixed Actor taxonomy. We use these fields only for aggregate reliability, repeat-use and cross-Actor analysis; reports suppress any cell with fewer than five distinct users.

The same disableUsageStats: true input flag turns these fields off too. The user ID is removed after 13 months; we do not export, sell, share, or attempt to re-identify this data.

Run-outcome signals (v2)

To learn whether a run did what it was asked to do, the record also holds a few more coarse ranges and yes/no flags. None of them contains content:

• the result limit you asked for (a range, when the input has one) and what share of it was delivered;
• results delivered per input item you listed (a range);
• output quality as ranges: how fully the result fields were filled, the share of rows that look like errors, the share of duplicate rows, and how many different fields appeared. These are counted in memory while results are saved; no result content is kept;
• how the run was started (console, API, schedule, webhook, another Actor);
• how it ended: stopped by you, timed out, reached the requested limit, stopped by the charge limit, and how many times the platform moved the run;
• if this Actor reports it: how many items to process worked or failed (ranges) and one failure reason from a fixed list;
• a short code made from the names of the input fields you set, never their values.

Repeat-run fingerprint (v2)

When your Apify user ID is recorded (see above), the record also holds an 8-character one-way code made from your input (proxy settings left out) and this Actor's name. It only lets us see that the same account ran the same input again soon after an unsatisfying run; we never see the input itself. It is stored only in the database, never published, and reports use it in aggregate with the same five-user minimum. It is the one exception to the statement above that no hashes are collected, and disableUsageStats: true turns it off.