SkillSpace

SkillSpace is a semantic resume-to-job matching project that uses compact embedding models, FAISS retrieval, and deterministic skill-gap reasoning to produce explainable career-match outputs.

Phase 1 is intentionally scoped as a deployable vertical slice rather than a from-scratch transformer research project.

---

Vision

Modern hiring pipelines rely heavily on keyword heuristics and static rule-based systems. SkillSpace explores a learning-based approach where:

• resumes are represented as semantic embeddings
• jobs are represented in the same vector space
• similarity reasoning replaces keyword matching

The goal is to build a career intelligence engine rather than a simple resume matcher.

---

Core Idea

SkillSpace maps resumes and jobs into a shared embedding space, retrieves relevant roles semantically, and explains the result with lightweight skill-gap reasoning.

Phase 1 uses a public encoder model:

• sentence-transformers/all-MiniLM-L6-v2

---

🏗️ System Overview

flowchart TD

User[User uploads resume]
UI[Gradio Interface]
Parser[Resume Text Processing]
Encoder[Dual Encoder Transformer]
EmbedDB[Vector Index / FAISS]
Jobs[Job Dataset]
Logic[Post Processing Engine]
Output[Career Insights UI]

User --> UI
UI --> Parser
Parser --> Encoder
Encoder --> EmbedDB
Jobs --> EmbedDB
EmbedDB --> Logic
Logic --> Output

Loading

---

Phase 1 Features

• Semantic job matching
• Skill gap recommendations
• Hugging Face dataset ingestion
• Retrieval evaluation with Recall@K / MRR
• Gradio demo with sample and HF runtime modes

---

Current Stack

• PyTorch
• Sentence Transformers
• FAISS
• Hugging Face Datasets / Models / Spaces
• Gradio

---

🗺️ Planning Docs

• Feasibility Assessment
• Phase 1 Implementation Plan
• Phase 2 Implementation Plan

---

Project Status

Current state:

• public HF datasets downloaded and normalized
• job embeddings generated with a compact encoder
• FAISS retrieval working
• Gradio app working in sample and hf modes
• evaluation script added
• public Hugging Face Space live

Current artifacts:

• processed jobs: artifacts/processed/jobs_hf_v1.parquet
• processed resumes: artifacts/processed/resumes_hf_v1.parquet
• embeddings: artifacts/embeddings/jobs_hf_v1.npy
• FAISS index: artifacts/index/jobs_hf_v1.faiss

Live links:

• Demo Space: greatvivek11/skillspace-demo
• Dataset repo: greatvivek11/skillspace-data
• Model repo: greatvivek11/skillspace-encoder

---

Local Run

Install dependencies:

python3 -m pip install -r requirements.txt

Run the app:

export SKILLSPACE_HF_LOCAL_ONLY=1
python3 -m src.ui.app

If HF artifacts are present, the app now defaults to hf mode automatically. Otherwise it falls back to sample.

Hugging Face Space Prep

Recommended publishing split:

• Space repo: Gradio app and lightweight runtime code
• Model repo: fine-tuned embedding model
• Dataset repo: cleaned SkillSpace dataset artifacts

Space-specific notes:

• root entrypoint: app.py
• Space dependencies: requirements-space.txt
• offline-friendly local cache mode supported via SKILLSPACE_HF_LOCAL_ONLY=1
• HF runtime prefers prebuilt index artifacts automatically when available

Portfolio Positioning

This Phase 1 release is best framed as:

• a semantic candidate-to-role matching demo
• an explainable retrieval system for recruiters and hiring managers
• a deployable ML product slice rather than a pure research prototype

Motivation

SkillSpace is built to explore how modern representation learning can reshape career intelligence systems.

The project aims to demonstrate practical ML system design, retrieval engineering, and product thinking in a portfolio-friendly form.