Job exploration
Search and filter a dataset of 1,658 ranked postings. Synonym handling broadens related-term matching; it is not embedding-based semantic search or a live job feed.
AI Job Market
Copilot
Explainable Job Matching, Skill Gap Analysis & Local AI Career Insights
A local-first career intelligence prototype that helps users explore a ranked set of job postings, understand recorded skill-match signals, and review requirements that still need verification. It combines a Python scoring workflow with contextual analysis generated locally through Ollama and Qwen3 4B.
Demo availability depends on the host laptop, Streamlit, Ollama, and ngrok staying online.
01 / Problem
Turning a long list of jobs into interpretable signals
A ranked list can help narrow opportunities, but a score alone does not explain which recorded skills overlap with a role, which areas deserve review, or what important requirements are missing from the available data. The project explores how to make those signals easier to inspect without presenting an AI-generated opinion as verified fact.
02 / What I built
A workflow from ranked data to contextual insight
Explainable fit signals
Display the rule-based fit score, data coverage, experience-fit estimate, recorded matched skills, and potential development areas for a selected role.
Contextual local AI
Use a locally running Qwen3 4B model via Ollama to connect recorded candidate skills to responsibilities in the stored job description and propose practice actions and hypothetical interview questions.
Data-quality awareness
Surface missing or uncertain requirements so users can distinguish recorded signals from conclusions that still require verification.
01 · DataRanked CSVExisting job postings and scoring fields.
02 · LogicPython + PandasFilter, ranking signals, coverage and skill flags.
03 · InterfaceStreamlitOverview, job exploration, match insight and quality review.
04 · GenerationOllama + Qwen3 4BContextual career insight kept separate from the official score.
03 / Snapshot
What the current dataset shows
1,658
Ranked job postings available in the current dataset.
13
Initial global shortlist count in the current application summary; this is a ranking output, not a hiring outcome.
651
Postings marked “Needs Review” in the current summary.
Local-first AI
Contextual insights are generated through an Ollama endpoint running on the user's machine; no paid LLM API key is required for that local workflow.
04 / Design decision
Keep scoring and generation separate
The rule-based fit score is not rewritten by the LLM. The scoring pipeline supplies the fit signal, matched-skill records, potential development areas, and known data limitations. The local model adds contextual explanation and suggestions, which still require human review.
05 / Validation
Representative scenarios reviewed
The local workflow was reviewed using two contrasting roles: Business Analyst at Purplle and Senior Data Analyst at Introhive. These examples exercise different responsibilities and seniority context. Offline grounding tests and a local QA script were also run. This is representative testing, not exhaustive validation of all 1,658 records or a benchmark of hiring outcomes.
06 / Limitations
What this prototype does not claim
- It uses an existing ranked dataset; it does not fetch live job postings.
- Role searches use job-title logic and synonym matching, not embedding-based semantic search.
- Experience estimates can be heuristic and are not necessarily explicit employer requirements.
- Education requirements may be unavailable, so eligibility cannot always be confirmed.
- Skill matches are scoring signals, not independent proof of proficiency.
- AI-generated content may still be wrong and should be checked against stored job-description evidence and the candidate's actual background.
- The application uses local Ollama; a hosted public demo is not provided by this page.
07 / Next steps
A foundation for further evaluation
The next improvements would be broader, documented validation across a varied sample of job descriptions, clearer evaluation criteria for contextual relevance, and a privacy-safe, licence-compliant source-code release. The current version prioritizes explainability and careful treatment of missing data over claiming automated hiring accuracy.