TalentSonar: From Job Descriptions to GitHub Candidates
A 24-hour hackathon prototype: Gemini parses a job description into structured requirements, and the GitHub GraphQL API finds real candidates who match.
2 people
Team
Hackathon prototype
Format
Gemini API
Job-ad parsing
GitHub GraphQL
Candidate search
Hypothesis
Can an LLM turn a free-text job description into a structured search that finds real candidates on GitHub, fast enough to be useful in a 24-hour hackathon prototype?
- Problem
- Turning a free-text job ad into a search for people is slow manual work for recruiters. The prototype tests whether an LLM plus the GitHub API can produce a first candidate list straight from the ad.
- My role
- 24-hour hackathon, team of 2 with Amirhosein Shahdadian. I built the job-ad parsing, the GitHub discovery and the Streamlit pages.
- Result
- Gemini turns a job ad into a typed requirement list, and async GitHub GraphQL searches return a ranked list of real public developers with stated reasons.
Honest limits
- Candidates are ranked by a simple rule-based score that was never validated, and years of experience are random placeholders, so the ranking is not a real assessment.
- LLM scoring of candidate answers was not built; text answers are scored by length.
From a job ad to a first list of candidates
Recruiters start from a job description written in free text, and turning it into a search for people takes time. TalentSonar tries to shorten that step: the job ad goes in, structured requirements come out, and those requirements drive a search over public GitHub profiles.
Four steps from a job description to a candidate list
- 1Parse
An HR user pastes or uploads a job description (PDF, Word, or markdown). Gemini turns it into structured requirements: technical skills, soft skills, and the experience asked for.
- 2Search
The required skills are mapped to programming languages and sent to the GitHub GraphQL search API, which returns real public GitHub users who match.
- 3Review
Candidates land in a Streamlit HR dashboard next to the job posting, where the recruiter can review profiles and track jobs, candidates, and invitations.
- 4Assess
The recruiter can send a candidate a questionnaire (multiple choice, text, and rating questions) in English or Italian.
What works and what does not
- Works: Gemini parsing of job descriptions into structured requirements, and GitHub GraphQL search that returns real candidate profiles.
- Placeholder: candidates are ranked by a simple rule-based score that was never validated, and years of experience are random placeholders, so the ranking is not a real assessment.
- Not built: LLM scoring of candidate answers; text answers are scored by length.
Frameworks and infrastructure
Source code on GitHub.