Job Searcher

Source: HuggingFace Date: 2026-06-06 URL: https://huggingface.co/blog/build-small-hackathon/job-search-blog

Summary

Job Searcher is a HuggingFace hackathon project that takes a resume and preferences (job type, location, modality) and returns a curated, scored shortlist of LinkedIn positions. The architecture is a teacher-student distillation pipeline: DeepSeek V4 Pro labeled a 10,000-job training dataset offline; Qwen3-8B was then fine-tuned with two separate LoRA adapters (rank 16) — one for query generation, one for resume-job fit scoring — hot-swapped during inference. LinkedIn search runs via JobSpy. The Space deploys on llama.cpp with CUDA on Hugging Face ZeroGPU, with OpenAI-compatible streaming output.

Implications

• The model-capability freeze. This is a clean example of capability unlocked through fine-tuning rather than frontier model access: an 8B model doing domain-specific scoring at quality that required a frontier teacher to label. The pattern (distill a large model’s judgment into a small model’s weights for a narrow task) is becoming standard infrastructure.
• Supply-chain trust. The system reads live LinkedIn data via scraping (JobSpy) — a source that is neither stable nor officially sanctioned. Any production deployment of this pattern inherits upstream fragility from the scraping layer, which is an unresolved supply-chain dependency hiding behind a clean inference stack.
• Token-cost economics. Two LoRA adapters hot-swapped on a single 8B base is an efficient inference design — the model serves two tasks without doubling the parameter count or inference cost. This is the kind of deployment-cost optimization that becomes meaningful at fleet scale.