Open Source AI Assistant Reduces Job Search Friction for New Graduates Developers from the build-small-hackathon community have released an AI-driven job search assistant designed to streamline the application process for recent graduates. Hosted on Hugging Face Spaces, the tool automates tedious recruitment phases by combining transparent reasoning with lightweight model architecture. The initiative addresses a market inefficiency where early-career professionals spend months manually filtering hundreds of postings, often submitting generic applications misaligned with their objectives. The assistant operates through a three-stage workflow. Initially, the model analyzes a user resume alongside configured preferences such as work modality, location, and role type to generate targeted search queries. These queries are executed against LinkedIn via a utility called JobSpy. The system then evaluates each posting by scoring the alignment between the resume and the job description across five distinct dimensions. Rather than returning an overwhelming list, the platform delivers a concise shortlist complete with token-by-token reasoning that explains ranking decisions. The underlying architecture employs a knowledge distillation strategy to balance performance with computational efficiency. The development team utilized DeepSeek V4 Pro as an offline teacher model to generate structured training labels across a closed-loop corpus. The distilled knowledge was transferred to Qwen3-8B, quantized to Q4_K_M to operate within a single ZeroGPU inference slice. Training was executed through two separate LoRA supervised fine-tuning runs on a single A100 GPU, isolating query generation from fit evaluation to prevent output formatting conflicts. Deployment utilizes llama-cpp-python on a Hugging Face ZeroGPU environment, streaming responses through an OpenAI-compatible endpoint to render reasoning tokens directly in the interface. The project emphasizes full transparency, publishing the complete Claude Code development session as a native trace dataset. Researchers can examine every iterative decision and debugging phase via the official Hugging Face agent-traces repository. Developers report two critical optimizations during the build process. Separating the query generation and evaluation tasks into dual LoRA adapters on the same base model eliminated cross-task formatting leakage. Furthermore, refining the teacher model scoring prompt to demand explicit resume to role comparisons significantly improved the student model alignment accuracy. The distillation process successfully transferred this granular evaluation habit to the smaller architecture. The assistant is currently available for public testing at the official Hugging Face Spaces repository. The project provides a replicable framework for automating preference-driven digital workflows while maintaining computational accessibility for individual developers.