brighttalent
AI-powered talent search and analysis platform built on top of the Avionté ATS, using vector embeddings and GPT-4 for semantic candidate matching.
46 commits over 8 days
What it is
Bright Talent integrates with the Avionté Applicant Tracking System to provide intelligent, natural-language talent search. Users can ask questions about their talent pool in plain English, and the system uses vector similarity search over candidate resumes, skills, and work history to surface relevant matches. It includes a chat interface for conversational queries, detailed candidate profiles with skills and experience, and an admin approval system to control access.
Why I built it
Built to modernize talent search in an existing ATS by layering on semantic search and AI-powered chat. The developer was likely exploring how to apply embeddings and LLMs to real-world recruiting workflows, solving the problem of keyword-based search missing semantically similar candidates. The multi-phase planning docs suggest this was a learning project in agentic software engineering patterns—incremental feature rollout, performance debugging, and complex integrations.
Tech stack
Next.js 14 (App Router), TypeScript, React; tRPC for type-safe API; Prisma ORM with PostgreSQL + pgvector extension for vector search; NextAuth.js with Google OAuth; OpenAI GPT-4 and Embeddings API; Tailwind CSS and Radix UI components. PDF/DOCX parsing for resume ingestion.
Timeline & effort
Last active Jul 24, 2025 — 46 commits across 8 active days (spanning 78 days).
Committed across 78 calendar days (May–July 2025) with 46 commits concentrated on 8 active days. Activity clustered heavily in July (29 commits), suggesting a couple of focused development sprints. Moderate sustained effort, likely part-time work over weekends or evenings.
Did it turn out useful?
The repo has a live Vercel deployment and substantial implemented features (vector search, chat, admin approval, resume parsing), but the detailed planning docs describe unfinished work (match reasoning not fully implemented, performance fixes for resume extraction documented but not resolved). The presence of multiple .md implementation plans, SQL debug scripts, and performance fix notes suggests active development still in progress. Evidence leans toward a working prototype with real capability but incomplete polish.