WRP HR Scorer: A Recruitment Scoring System Based on Local LLM

WRP HR Scorer is a fully locally-run AI recruitment scoring service implemented based on Ollama and the Mistral 7B model, enabling multi-dimensional resume evaluation without calling cloud APIs. Addressing data privacy concerns in HR scenarios, the project uses a layered architecture to separate LLM (a non-deterministic component) from business logic (deterministic Python code), balancing the utilization of AI capabilities with system testability and interpretability. Core features include eligibility review (6-month restriction on repeated applications), multi-dimensional scoring (qualifications/experience/achievements/cultural fit), robust error handling, and batch processing, providing enterprises with a privacy-friendly and maintainable recruitment assistance solution.

With the maturity of LLM technology, enterprises are exploring AI applications in HR processes, but sending sensitive resume data to cloud APIs poses compliance risks. WRP HR Scorer is a localized solution proposed in this context; by running the Mistral 7B model locally via Ollama, it avoids cloud calls and addresses data privacy issues. The project's design philosophy emphasizes LLM as the only non-deterministic component, while business logic remains in deterministic code, balancing AI capabilities and system reliability.

The system adopts a modular pipeline architecture with clear component responsibilities: CLI entry → PDF extraction → eligibility review → LLM scoring → score adjustment → persistent output. The core principle is that LLM only returns raw scores for four dimensions, while business logic such as recommendation bonuses and grade classification is handled by deterministic Python code. This design brings three advantages: 1. Business rules can be unit-tested; 2. The model only accesses resume content without involving sensitive business logic; 3. The final results are supported by clear rules, ensuring interpretability.

Core features include: 1. Multi-dimensional scoring: LLM scores resumes from four dimensions (0-100 points: qualifications, experience, achievements, cultural fit), and the system maps these to A/B/C/Reject grades after applying internal recommendation bonuses; 2. Eligibility review: Blocks candidates rejected within 6 months from re-applying and accurately calculates the earliest re-application date; 3. Robust error handling: Model outputs undergo parsing, validation, retries (with strict prompts), and safe fallback (zero score → Reject) to ensure system stability; 4. Batch processing: Implements batch scoring of resumes in a folder and generates rankings via batch.py, reusing core pipeline functions.

The project chooses Ollama as the local LLM runtime environment, with advantages including simplified model management (download via 'ollama pull'), OpenAI-compatible API, and high resource efficiency (Mistral7B runs on consumer-grade hardware). The tech stack also includes pdfplumber (PDF extraction), SQLAlchemy (data persistence), and Pydantic (data validation)—concise, mature, and without redundancy.

Current limitations and improvement areas: 1. Candidate ID relies on CLI parameters or file names and is not parsed from CV text, requiring additional data preparation; 2. Code organization needs optimization (e.g., functions imported in batch.py should be extracted to utils.py); 3. Dependent on Mistral7B's instruction-following ability; unstable output formats require prompt adjustments or model replacement; 4. Scoring dimensions are hard-coded, needing support for customization across different industries/positions.

WRP HR Scorer's insights for the HR Tech field: 1. Privacy-first design eliminates data leakage risks; 2. Human-machine collaboration (AI assists decision-making, humans make final judgments); 3. Auditability (persistent application records form an audit trail); 4. Progressive automation (expanding capabilities from single CV to batch processing). The project proves that consumer-grade hardware can build valuable localized enterprise AI solutions, providing a reference implementation for teams concerned about cloud risks.