NLP and entity extraction
Worked with spaCy and transformer-based approaches, including a fine-tuned DeBERTa-v3 model, to extract structured information from resume text.
An AI-assisted recruitment workflow designed to turn unstructured resumes into structured candidate information and support screening, matching, and recruiter review.
Problem space
Recruitment automation
AI focus
NLP + transformer models
Workflow
Extract → Match → Review
01 / Context
Recruiters receive resumes in many different formats. Reviewing each document, extracting candidate details, and comparing them with job requirements can require substantial manual effort.
The platform brings several stages of this workflow together: resume ingestion, structured information extraction, candidate matching, search, and recruitment workflow support.
The goal is decision support for recruiters — not an automated replacement for human evaluation.
02 / Solution
Document intake
PDF and DOCX resume uploads
Text & entities
NLP and transformer extraction
Candidate profile
Structured information for review
Matching
Semantic similarity with job criteria
Recruiter workflow
Search, screen and review
Conceptual workflow for public portfolio presentation. It is intentionally simplified and does not expose private employer architecture or data.
03 / My contribution
Worked with spaCy and transformer-based approaches, including a fine-tuned DeBERTa-v3 model, to extract structured information from resume text.
Used all-mpnet-base-v2 sentence embeddings and cosine similarity to compare resume content with job-description content.
Integrated language-model workflows for resume summarization, categorization, and structured experience extraction.
Worked across Django REST Framework and FastAPI services, with Celery and RabbitMQ for background processing and Elasticsearch for search.
04 / Challenges
Resumes vary in layout, wording, section order, and completeness. The extraction workflow needs to handle this variability and represent the result in a consistent structure.
Transformer-based entity extraction and NLP processing are used to identify candidate information. Model outputs still require validation and careful handling of missing or ambiguous fields.
Resume and job-description text can express similar skills in different language. Semantic representations provide a complementary signal to conventional keyword-based matching.
Large language models can place significant pressure on GPU memory. Controlled queues and concurrency limits help avoid overloading a single inference device.
05 / Technology
06 / Evaluation
Extraction quality
Entity-level precision, recall, and F1 across fields such as names, skills, education, and experience.
Matching relevance
Human-reviewed relevance, ranking quality, and error analysis for different job families.
Operational reliability
Batch completion rate, processing latency, queue behavior, resource utilization, and failure recovery.
Measured accuracy, throughput, and latency figures are not published here because verified, reproducible public measurements have not been supplied for this case study.
07 / Reflection
Building AI features is only one part of delivering an AI product. Input variability, model behavior, API design, asynchronous execution, infrastructure limits, and user trust all matter. This work strengthened my interest in applied AI engineering: selecting and integrating suitable models, evaluating their outputs, and connecting them to dependable applications. Backend systems, asynchronous workflows, and deployment are important foundations that help make those AI capabilities useful in practice.
Want to explore more of my work?
See the other projects and interactive demos.