DEV Community

Leul Yohannes
Leul Yohannes

Posted on

I built a skill verification pipeline for AI engineers — here's how

The Problem

100,000+ people "learn AI" every month, but less than 10% can ship production code.

Employers can't tell who can build vs who watched tutorials.

The Solution

I built Forge AI — a skill verification pipeline that replaces courses with real engineering work.

Here's how it works:

1. Real Engineering Tickets

Users get 5 real RAG tickets from a fictional company (Meridian AI):

  • RAG-114: Chunking returns empty results on long documents
  • RAG-115: Hybrid score fusion produces incorrect rankings
  • RAG-116: Metadata filter ignores date conditions
  • RAG-117: Reranker returns raw scores instead of final documents
  • RAG-118: Faithfulness evaluator always returns 0%

2. Write Code in a Real Editor

The workspace has:

  • A task description panel (the ticket)
  • A Monaco code editor with starter code
  • A terminal panel that shows results

3. Automated Evaluation

When a user submits code:

  1. It runs in a sandbox (timeout + memory cap)
  2. Pytest suites check correctness
  3. An AI judge (Groq LLM) provides detailed feedback

4. Verified Skill Profile

Users get a talent graph showing scores across 5 competencies:

  • Chunking
  • Fusion
  • Filtering
  • Reranking
  • Evaluation

The Tech Stack

  • Frontend: Next.js 16.2.10, React 19, TypeScript, Tailwind CSS 4
  • Editor: @monaco-editor/react
  • Backend: Python 3.13, FastAPI
  • Database: Supabase (PostgreSQL)
  • Auth: Supabase Auth
  • AI Judge: Groq API (llama-3.3-70b-versatile)
  • Hosting: Vercel (frontend), Render (backend)

The Architecture

The backend pipeline is a plain Python function composition:

Top comments (0)