TalentSonar: Developer-Skill Inference from GitHub

Every technical recruiter eventually faces the same problem: a candidate's GitHub profile is the most honest signal of their actual coding, but it's also dozens of repos of varying quality, mixed languages, and side projects vs serious work. Reading a profile cold takes 20 minutes per candidate.

TalentSonar collapses that into a single page: paste a handle, get an LLM-grounded read on technical skills, project archetypes, and seniority signal, with the underlying evidence (repos, commit cadence, languages) visible alongside.

1. 1Pull

   GitHub GraphQL API extracts public-repo metadata, commit history, and language breakdowns for any handle in a single round trip.

2. 2Structure

   Raw API output is summarised into compact features: tech stack distribution, project archetypes (CRUD app vs ML pipeline vs library), commit cadence.

3. 3Infer

   Structured features are passed to Google Gemini with a prompt that asks for skill inference, seniority signal, and stand-out projects, with grounded explanations per claim.

4. 4Score

   A Streamlit candidate-scoring view combines the LLM read with a configurable test layer. PDF report downloadable for downstream review.

• Live today: GitHub GraphQL extraction, Gemini-driven inference, Streamlit candidate-scoring UI, CSS-only anti-cheat layer, downloadable PDF report.
• Mock today: the candidate test scoring is a heuristic placeholder; the next iteration swaps it for LLM-evaluated scoring with explanations per question.
• v2 plan: a Recruiter dashboard with multi-candidate sorting, JD-match scoring, and a link to schedule directly with the strongest match.

The Streamlit app lives on HuggingFace Spaces. Paste a GitHub handle to see the GraphQL extraction plus the Gemini-driven skill inference end to end.

Source code on GitHub.