Saevix Software Agency — CareerByAI
Parsed and structured 50,000+ unstructured resumes into actionable learning trees
The Problem
Youth career counseling often relies on static, generic advice that fails to adapt to modern tech roles. CareerByAI needed a system to ingest wildly unformatted PDF resumes from students, extract their genuine skills without hallucination, and programmatically match them against real-world job requirements in alignment with UN Sustainable Development Goal 8.
The Approach
We architected an ingestion pipeline utilizing PDF.js on the client to extract text buffers before passing them to the Node.js backend, reducing server bandwidth. To prevent AI hallucinations during skill extraction, we engineered strict few-shot prompts using the Google Gemini AI API, forcing the output into a deterministic JSON schema. We rejected heavy vector databases like Pinecone for this MVP, instead utilizing MongoDB's native aggregation framework to perform semantic-like matching between the structured resume JSON and our curated course database.
The Result
The platform successfully parsed over 50,000 resumes with an 89% structural accuracy rate, generating visual roadmap trees instantly. By moving the PDF text extraction to the client, we reduced backend processing overhead by roughly 30%.
Lessons
Relying on LLMs strictly for JSON generation is powerful but fragile. If we were to scale this further, we would implement a dedicated NLP pipeline (like spaCy) for the initial entity extraction, using the LLM exclusively for the complex matching logic rather than raw data parsing.
Results
Roadmaps Generated
Parsing Accuracy
Server Load