The End of the Infinite Scroll: Precision Job Matching via Distillation

The era of scrolling through endless, irrelevant job boards is being replaced by structured, reasoning-based filtering. The Job Searcher framework demonstrates how small, specialized models can outperform brute-force searching by applying high-level reasoning to candidate data.

The system operates through a three-step pipeline: generating LinkedIn-shaped search queries, executing those queries via JobSpy, and scoring the resulting postings. Rather than returning a massive list of roles, the framework produces a small shortlist accompanied by defensible reasoning across five dimensions: skills match, experience relevance, education and certifications, industry/domain fit, and seniority alignment.

The technical architecture relies on a teacher-student distillation method. The teacher, DeepSeek V4 Pro, acts as a label generator to create structured judgment. The student, Qwen3-8B, is designed to absorb this judgment. This student model is small enough to fit on a single ZeroGPU slice once quantized to Q4_K_M.

The training corpus was built through a closed-loop process using 2,500 resumes from the Divyaamith/Kaggle-Resume dataset. The teacher drafted queries for each resume, which were then used to scrape approximately 10,000 postings via JobSpy. The teacher subsequently scored every (resume, job) pair across the five dimensions, providing one sentence of reasoning per dimension.

The training process involved two LoRA SFT runs on a single A100 via Modal. The configuration utilized a rank of 16 and an alpha of 16. To ensure stability, the schedule included mid-epoch checkpoints every 200 steps.

This approach signals a shift in developer tooling from general-purpose LLM prompting to highly specialized, distilled agents. When a model can explain why the second-ranked job beats the third, the value moves from the quantity of data to the quality of the reasoning.

The question for developers is no longer how to scale models, but how to effectively distill complex, multi-dimensional reasoning into models small enough to run on edge infrastructure.

Source