The quest for a first job after graduation often feels like a full-time occupation in itself. You dedicate countless hours, perhaps even months, to sifting through an endless stream of job postings, hoping to find just a handful that are truly worth your time. The repetitive cycle of clicking “Easy Apply” and crafting custom cover letters for dozens of roles can quickly lead to burnout, pushing you towards applying for positions you don’t genuinely want.
This relentless grind often results in a state of exhaustion, where the mental effort of evaluating each job listing becomes more taxing than simply submitting an application, regardless of fit. You end up applying to roles in industries you barely care about, just to feel productive. We understood this challenge deeply, recognizing that new graduates needed a smarter, more efficient way to navigate the turbulent waters of the job market.
That’s precisely why we developed an innovative AI-powered job search assistant designed to transform this arduous process. Imagine dropping your resume, then watching as our system intelligently processes countless opportunities, presenting you with a highly curated shortlist. It’s about replacing overwhelming volume with precise, relevant matches, saving you invaluable time and energy.
Streamlining Your Job Search: A Smarter Approach
Our job search assistant doesn’t just give you a generic list; it provides a carefully selected shortlist with transparent, defensible reasoning behind each recommendation. You can clearly understand why the model believes one job is a better fit than another. This level of insight empowers you to make informed decisions, focusing your efforts on roles where you truly have a strong chance.
The tool cuts through the noise, identifying positions that align not just with keywords, but with the nuances of your experience and skills. It transforms a broad, draining search into a focused, strategic endeavor. This means less time sifting through irrelevant postings and more time preparing for interviews that truly matter to your career trajectory.
We believe that your time is too valuable to waste on endless, unfocused applications. By offering a concise list accompanied by detailed explanations, our AI job search assistant acts as your personal career strategist. It ensures every application you send out is a high-quality, targeted effort, maximizing your chances of landing that dream role.
Behind the Scenes: The Tech That Powers It
To build this intelligent system, we adopted a sophisticated “teacher-student” model architecture, blending the strengths of advanced language models. The “teacher” in our setup is DeepSeek V4 Pro, a powerful model known for its structured reasoning capabilities and strict adherence to output schemas. We utilized it offline to generate high-quality labels over a vast corpus of job data, essentially teaching our system what constitutes a good job match.
Our “student” model is Qwen3-8B, chosen for its efficiency and robust ability to absorb complex judgments from the teacher. This smaller, yet highly capable, model is optimized to run on accessible hardware, specifically a single ZeroGPU slice when quantized to Q4_K_M. This ensures that the powerful insights gleaned from the teacher can be delivered quickly and cost-effectively to users in real-time.
The training corpus itself was meticulously curated from a closed-loop, resume-aware end-to-end process, ensuring high relevance and accuracy. For training, we employed two LoRA (Low-Rank Adaptation) SFT (Supervised Fine-Tuning) runs on a single A100 via Modal, one for each specific task. This specialized training approach allowed us to fine-tune the model effectively.
The LoRA configuration was specifically designed to enhance performance across multiple crucial components of the model:
r=16,lora_alpha=16: Key parameters for LoRA adapter size and scaling.task_type="CAUSAL_LM": Indicating the model’s primary language generation task.target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]: Targeting critical attention and feed-forward network layers for optimal adaptation.
Experience the Difference: Live Demo & Development Insights
The live inference space for our job search assistant runs on a HuggingFace ZeroGPU Space, powered by llama-cpp-python with a pre-built CUDA wheel. A crucial design choice was to enable streaming output, utilizing the OpenAI-shaped create_chat_completion(stream=True) function. This ensures that the detailed reasoning for each job match lands in the user interface token by token, providing an engaging and immediate experience.
For those interested in the nitty-gritty of development, we’ve published the entire Claude Code session that built this space as a HuggingFace agent-traces dataset. This includes raw JSONL events and offers a transparent look into every decision, dead end, and recovery made during the development process. It’s a goldmine for anyone curious about the actual iterative process behind building such a tool.
During development, we uncovered several valuable insights. Notably, employing two separate LoRA adapters proved significantly more effective than a single one. Initially, we attempted to combine query generation and fit evaluation into one adapter, which led to frustrating formatting leaks—JSON bleeding into prose and vice versa. Splitting these into two distinct heads, hot-swapped per call, completely eliminated these bugs, showcasing the power of specialized adaptation.
Another profound learning was the immense impact of the teacher’s prompt quality on the student model’s output. By refining the teacher model’s labeling prompt to score against specific resume details, such as “four years of Rust; the role asks for five,” instead of vague “strong technical match,” the student model naturally adopted this granular, detailed reasoning. This direct propagation of high-quality instruction through distillation significantly enhanced the student’s analytical capabilities.
Ready to transform your job hunt? Stop the endless sifting and let our AI job search assistant do the heavy lifting for you. Visit huggingface.co/spaces/build-small-hackathon/job-search-assistant today, drop your resume, and discover a smarter way to find your next career opportunity.
Source: Hugging Face Blog
