Enhancing ERP with LLM-Based Salary Parsing: Evrone’s Engineering Approach

AI adoption in HR is accelerating, but real-world implementation often reveals unexpected complexity. Evrone approached AI integration not as hype, but as engineering discipline.

The Context

Evrone operates a custom ERP covering the entire recruitment lifecycle. Yet recruiters at Evrone spent hours normalizing CV data:

• Copying salary expectations
• Converting currencies
• Interpreting vague phrases
• Aligning formats with ERP requirements

Regex patterns were brittle. Generic LLM prompts hallucinated. Evrone needed precision.

Architecture Overview

Evrone implemented a two-stage LLM pipeline:

1️⃣ Resume Parsing with Qwen

Evrone configured Qwen for structured output via JSON Schema. The model extracts:

• Position category (backend, frontend, QA, etc.)
• Grade
• English proficiency
• Location
• Salary expectations

Structured outputs eliminated post-processing complexity.

2️⃣ Salary Standardization with YandexGPT

Evrone collected 10,000 historical salary records from Huntflow.
The team:

1. Performed baseline inference
2. Annotated inconsistencies
3. Fine-tuned YandexGPT 5 Lite with LoRA

The result:

• 95% parsing accuracy
• 97% correct USD recognition
• Minimal hallucinations

Impact

Within three months, Evrone reduced external HR system queries by 90%. Recruiters now review only edge cases (≈5%). Sales teams access reliable salary data instantly.

Key Insight

Evrone demonstrates that AI success depends on:

1. Narrow scope
2. High-quality training data
3. Clear structured outputs
4. Human-in-the-loop validation

Instead of replacing recruiters, Evrone augmented them. And that distinction makes all the difference.

From Regex to LoRA: How Evrone Optimized HR Data with LLMs.