AI salary parsing: from $180k-DOE strings to structured ranges

Salary is the field developers ask about most and the one that’s hardest to get right. Most ATSs don’t expose a structured comp field at all - the salary range, when it exists, is embedded in the job description as free text. And it’s embedded in roughly 11 different shapes.

Eleven shapes of “$180k”

Here’s a sample of the phrasings we’ve found in the wild:

• “$180,000 - $240,000”
• “$180K to $240K base + equity”
• “Compensation: $180,000-240,000 OTE”
• “Salary range: 180-240k USD”
• “Up to $240k DOE”
• “€180k cash + significant equity”
• “Comp band: $180k-$240k. DOE.”
• “$15,000-$20,000 per month”
• “DOE” (just that)
• “Pay for this role: 180000-240000”
• “Annual compensation: 180-240 thousand”

The parser, in three layers

Layer one is a regex matcher tuned to the 80% case. If a description contains $Xk-$Yk or $X,000-$Y,000 we capture it deterministically. This catches roughly 70% of postings.

Layer two is a structured prompt sent to a small language model - we use Haiku for cost. We pass the description and ask for min, max, currency, period, and includes_equity as a strict JSON schema. If the model returns confidence above 0.85 we accept it.

Layer three is the abstain layer. If neither the regex nor the LLM gives high-confidence output, we set compensation: null. A null is better than a wrong number - wrong numbers compound when customers filter by them.

What the output looks like

{
  "compensation": {
    "min": 180000,
    "max": 240000,
    "currency": "USD",
    "period": "yearly",
    "includes_equity": true
  }
}

Same shape, every source, every job. The free-text mess stays in the description field for anyone who wants it. Most don’t.