dobrandoCulleres

An AI platform using vector embeddings to match experts by skills, not job titles. It features an AI chatbot, smart team organization, and organization visualization tools.

Comment

Inspiration

The Problem: As organizations grow, talent tends to become invisible. When a manager needs someone with specific expertise, they don't have a way to find them beyond asking colleagues or guessing based on job titles. A developer who knows Kubernetes might be hidden in the ops team. A designer with machine learning experience might be unknown to the AI team.

Why It Matters:

  • Teams struggle to form because expertise is undiscovered
  • Projects get delayed waiting for the "right person"
  • Cross-functional collaboration is harder than it needs to be
  • Employee skills go underutilized
  • New hires can't find mentors or collaborators with shared interests

Our Solution: dobrandoCulleres (Galician for "bending spoons") is an intelligent talent discovery platform that makes invisible expertise visible. We use AI-powered semantic search to help organizations instantly find the right people for any project.

What It Does

Core Features

1. Intelligent Skill Discovery

  • Semantic skill search using vector embeddings (Ollama + pgvector)
  • Find experts even when you don't know the exact skill name
  • Search "golang" and find people with "Go", "Go programming", "Golang", etc.
  • Relevance-ranked results by skill level and match score

2. Team Intelligence

  • Organize employees into groups/teams with designated leaders
  • View organizational hierarchy with interactive org charts
  • See team composition and skill distribution
  • Track which skills exist in which teams

3. Expert Finder

  • Search for experts across multiple dimensions:
    • By skill keywords (semantic matching)
    • By specific skills with level filters
    • By department or team
    • By skill level (beginner to expert)
  • Get relevance-scored results showing match confidence

4. AI-Powered Chatbot

  • Chat interface to ask for team recommendations
  • Natural language queries: "Who can help with Kubernetes?"
  • Get team suggestions for specific project needs
  • Discover expertise patterns in your organization

5. User Management & Admin Tools

  • Comprehensive user profiles with job titles and departments
  • Skill management with proficiency levels (1-5)
  • User roles (Admin, Manager, Employee) with role-based access control

6. Multi-Language Support

  • Built-in internationalization
  • Language-specific routing
  • Support for multiple locales

How We Built It

Technology Stack

Frontend: Modern React Ecosystem

- Next.js 
- React 
- TypeScript 
- Tailwind CSS 
- Radix UI (accessible components)
- React Query 
- React Hook Form + Zod (form validation)
- Framer Motion 
- XY Flow (interactive diagrams for org charts)

Backend: Go + PostgreSQL

- Go
- PostgreSQL
- pgvector (vector embeddings)
- Ollama (local embedding generation)
- Gorilla Sessions

Architecture Decisions

Why Go for the Backend?

  • High-performance API suitable for real-time search
  • Excellent concurrency handling
  • Easy deployment (single binary)
  • Built-in HTTP support

Why Vector Embeddings?

  • Semantic search works beyond keyword matching
  • Users can search in natural language
  • Skill similarity detection (Kubernetes ≈ K8s ≈ container orchestration)
  • Scalable with pgvector (native PostgreSQL support)

Why Next.js?

  • Server-side rendering
  • API routes for backend integration
  • Built-in optimization (image, font, code splitting)
  • Natural i18n support

Key Implementation Details

Skill Matching Pipeline:

  1. User creates/searches for a skill
  2. Text is embedded using Ollama's mxbai-embed-large model
  3. Embedding is stored in PostgreSQL pgvector column
  4. Vector similarity search finds related skills
  5. Results ranked by cosine similarity

Authentication & Authorization:

  • Cookie-based sessions with 30-minute timeout
  • Role hierarchy: Admin → Manager → Employee
  • Middleware enforces permissions at API level
  • Session keys configured via environment variables

Challenges We Ran Into

1. Vector Embedding Complexity

Implementing semantic search with vector embeddings was new territory for most of our team. Solution:

  • Studied pgvector documentation thoroughly
  • Tested embedding quality with various models

2. UI/UX for Complex Features

Making skill search and team recommendations intuitive for non-technical users. Solution:

  • Used familiar patterns (search bars, filters)
  • Clear labeling and help text
  • Interactive org charts for visualization
  • Chatbot as natural language interface

Accomplishments We're Proud Of

1. Intelligent Semantic Search

  • Implemented vector embeddings from scratch
  • Skill matching that understands synonyms and similar concepts
  • Fast search even with large datasets (pgvector is optimized)
  • Relevance ranking by match score

2. AI Integration

  • Chatbot feature that understands natural language queries
  • Demonstrates practical AI application (not just a demo)

3. Multi-Language Support

  • Built internationalization from day one
  • Language-aware routing

What We Learned

Technical Learnings

  1. Vector Embeddings Are Powerful

    • They enable semantic search that keyword-matching can't do
    • The quality of embeddings matters a lot
    • pgvector brings vector search to PostgreSQL efficiently

  2. Go is Perfect for APIs

    • Incredibly fast to build REST APIs
    • Deployment is simple (single binary)

Built With

Share this project:

Updates