Recruitment Automation: Killing the Manual Salary Transfer

Everyone talks about AI in HR, but getting it to actually work without weird mistakes is harder than it looks. We took one concrete problem and solved it.

The problem

Our ERP pulls data from Huntflow and job boards. Recruiters were losing 2–3 hours daily just moving salary numbers around. Open a resume, hunt for the digits, jump to another system, convert currency, standardize the format, paste it in. And if the data was off, finding the right person for a budget was pure guesswork.
Plus candidates write their expectations in a hundred different ways: "120+ net", "$5k-7k", "looking for 180+", "min 150", or just "open to your range". All had to be normalized by hand.

The fix

Template parsers died immediately – resumes are too messy. We used Qwen with structured output. Feed it a raw resume, get back clean fields: role category, timezone, English level, grade, location, salary.
Huntflow was worse. The text there is full of junk like "200/year" or "$3000". Regex was useless, base LLMs kept messing up. So we grabbed a dataset from old records, annotated it, and fine-tuned YandexGPT 5 Lite with LoRA – trained specifically to get amount and currency right.

The outcome

Huntflow lookups dropped 90%. Data accuracy hit 95%. Recruiters now manually check only 5% of cases – the ones with truly bizarre formatting. Time that used to vanish into copy-paste hell now goes into actual recruiting.

Want the same for your team? We're here to help. Full case here.