01 · Problem The Challenge
Job seekers frequently struggle to articulate their professional experience in a way that is impactful enough to pass Applicant Tracking Systems (ATS). While cloud-based AI tools exist, they often generate generic "AI-speak" and require sending personal data over the internet. The goal was to build a local-first, highly specialized application that could run entirely on a Mac's unified memory, transforming amateur text into metric-driven statements suitable for tech recruiting, while mathematically proving the improvement.
02 · Solution System Architecture
I engineered a Flask-based web application with a specialized Qwen/Qwen2.5-1.5B-Instruct model at its core. Instead of relying on generic base model capabilities, I injected a custom Parameter-Efficient Fine-Tuning (PEFT) LoRA adapter, fine-tuned on a proprietary synthetic dataset of tech resumes. To ensure speed, the weights are permanently fused upon boot, eliminating abstraction layers and utilizing PyTorch's Metal Performance Shaders (MPS) for Mac GPU acceleration.
To quantify the text transformation, I developed a rigid 100-point Heuristic Scoring Engine. This algorithm evaluates Flesch Reading Ease (penalizing overly complex or conversational text), enforces strict grammar rules via a local Java-based LanguageTool server, analyzes sentiment confidence using a HuggingFace pipeline, and measures action verb density against a proprietary keyword list.
03 · Engineering Technical Highlights
LoRA Fine-Tuning
Locally trained a custom LoRA adapter (r=16, alpha=32) targeting attention layers to transform the massive 1.5B parameter model into an "Expert Tech Recruiter" persona.
Hardware Optimization
Leveraged Apple Silicon unified memory by offloading tensor computations to the GPU via MPS, and reduced inference latency to under 20 seconds using early-stopping tokens.
100-Point Scoring
Engineered a multi-layered NLP scoring algorithm combining readability metrics, strict grammar checks, and proprietary keyword matching to evaluate CV quality.
04 · Results The Outcome
The AI CV Optimizer successfully rewrites amateur resume bullet points into highly professional statements. By running entirely locally with customized fine-tuning and hardware acceleration, it provides a fast, private, and mathematically proven tool for job seekers aiming for tech roles.