SkillGraph

Turn a resume and job description into a prerequisite-aware learning plan.

SkillGraph extracts skills from both documents, estimates current mastery, finds gaps, and orders those gaps against a dependency graph. It then attaches learning resources from a curated catalog.

Pipeline

Stage	Implementation	Purpose
Skill extraction	all-MiniLM-L6-v2	Match resume and job-description text to a predefined skill list
Optional fallback	Gemini 1.5 Flash	Classify sparse inputs against the same predefined list
Mastery scoring	Weighted rules	Combine frequency, recency, and job relevance
Dependency graph	NetworkX DAG	Identify missing prerequisites and valid next skills
Structural feature	Node2Vec embeddings	Represent each skill's position in the graph
Gap ranking	LightGBM LambdaRank	Rank candidate skills using job weight, mastery, graph score, and degree
Path construction	Greedy frontier traversal	Choose the highest-ranked skill whose prerequisites are met
Resource selection	Curated catalog lookup	Match each skill to a course or a direct learning-resource link

The local semantic matcher is the default. Gemini is used only when GEMINI_API_KEY is set and SKILLGRAPH_ENABLE_GEMINI=1.

Ranking model and evaluation

The committed LambdaRank dataset contains:

• 1,000 synthetic profiles, split evenly across software engineering and data science
• 19,756 skill-label pairs
• Five features: job-description importance, Node2Vec score, mastery, in-degree, and out-degree

The checked-in training summary records NDCG@5 of 0.7913 and NDCG@10 of 0.8532 on the synthetic holdout split. These scores measure ranking agreement with paths generated by the graph algorithm. They do not measure learning outcomes or performance on real users.

Training inputs, scripts, and artifacts are committed for inspection:

Artifact	Location
Synthetic data	backend/training/synthetic_data.json
Training pipeline	backend/training/
LambdaRank model	backend/models/ranker.lgb
Recorded metrics	backend/models/training_summary.md

Scoring logic

mastery(skill) = 0.31 * frequency + 0.336 * recency + 0.354 * job_match
gap = skill required by the job description and mastery(skill) < 0.6
priority(skill) = LightGBM(
job_importance,
node2vec_score,
mastery,
in_degree,
out_degree
)
path = repeatedly choose the highest-priority skill whose prerequisites are satisfied

Data and limits

• The repository includes 45 curated courses. Domain filtering exposes 24 for software engineering and 22 for data science, including shared entries.
• Skill extraction is constrained to the repository's domain lists. Unlisted skills may be missed or mapped to a nearby listed skill.
• Course matching is deterministic, but a useful resource link is not evidence that the learning path itself is correct.
• The ranking evaluation uses synthetic labels produced by the graph logic, not independent human labels.

The skill taxonomy and graph data reference the O*NET Occupational Database. The repository also includes links to the resume and job-description datasets used during development.

Quick start with Docker

docker compose up -d

Open the frontend at http://localhost:3000 and the API docs at http://localhost:8000/docs.

Local development

Backend

cd backend
pip install -r requirements.txt
uvicorn app.main:app --host 127.0.0.1 --port 8000 --reload

Frontend

cd frontend
npm install
npm run dev

Rebuild the model artifacts

bash backend/training/run_all_training.sh

On macOS, LightGBM may also require OpenMP:

export DYLD_LIBRARY_PATH="/opt/homebrew/opt/libomp/lib:$DYLD_LIBRARY_PATH"