Talentier

AI recruiting engine that turns job descriptions into structured rubrics, sources candidates, analyzes real code signals, runs multi-agent scoring, and delivers an evidence-backed, ranked shortlist.

Comment

Inspiration

Hiring pipelines are noisy, slow, and biased toward keyword matching. We wanted a system that can discover strong technical talent from real OSS contribution history, not just resume claims.

What it does

Talentier finds engineering candidates by combining two first-class sourcing layers: GitHub contributor intelligence and LinkedIn profile discovery. The system analyzes OSS contribution signals, surfaces ranked candidates, and links them with LinkedIn profiles where relevant so recruiters can evaluate both technical evidence and professional context side by side.

How we built it

  • Backend: FastAPI + LangGraph orchestration pipeline.
  • Data sources: GitHub repository and contributor APIs via modular tools.
  • Scoring: Multi-stage ranking using contribution activity, repo breadth, and optional code-quality analysis.
  • Frontend: React + TypeScript UI for search inputs, live pipeline status, and ranked candidate results.
  • System design: Source query → source scoring → identity resolution → aggregation → ranking.

Challenges we ran into

  • Search engine/LinkedIn scraping instability, blocking, and anti-bot challenges.
  • Long pipeline latency when code-quality LLM scoring was enabled at scale.
  • Duplicate frontend keys caused by repeated load-more results.
  • CORS-looking browser errors that were actually backend 500 failures.

Accomplishments that we're proud of

  • Reframed the product from brittle scraping to a GitHub-first OSS intelligence workflow.
  • Built a working multi-stage candidate pipeline with clear ranking logic.
  • Added safer profile linking behavior (GitHub primary, LinkedIn only on match).
  • Improved reliability with faster defaults and better error handling.

What we learned

  • Reliability beats breadth for talent sourcing products.
  • “CORS errors” often hide server exceptions; backend observability matters.
  • Identity resolution and deduplication are core, not optional.
  • LLM scoring is useful, but must be configurable to avoid performance bottlenecks.

What's next for Talentier

  • Add per-request toggles for deep scoring vs fast mode.
  • Improve identity matching confidence and evidence transparency in UI.
  • Add robust pagination and caching for larger candidate sets.
  • Introduce recruiter-facing shortlist workflows and export/reporting.
  • Expand source coverage (e.g., Stack Overflow/Kaggle) while preserving auditability.

Built With

Share this project:

Updates