The latest articles on DEV Community by Leul Yohannes (@leulyohannes). https://dev.to/leulyohannes https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4029970%2Faca6ea88-85ca-46d2-9f52-455b195fd275.jpg https://dev.to/leulyohannes en Leul Yohannes Wed, 15 Jul 2026 08:50:46 +0000 https://dev.to/leulyohannes/i-built-a-skill-verification-pipeline-for-ai-engineers-heres-how-26j4 https://dev.to/leulyohannes/i-built-a-skill-verification-pipeline-for-ai-engineers-heres-how-26j4 <h2> The Problem </h2> <p>100,000+ people "learn AI" every month, but less than 10% can ship production code.</p> <p>Employers can't tell who can build vs who watched tutorials.</p> <h2> The Solution </h2> <p>I built Forge AI — a skill verification pipeline that replaces courses with real engineering work.</p> <p>Here's how it works:</p> <h3> 1. Real Engineering Tickets </h3> <p>Users get 5 real RAG tickets from a fictional company (Meridian AI):</p> <ul> <li> <strong>RAG-114:</strong> Chunking returns empty results on long documents</li> <li> <strong>RAG-115:</strong> Hybrid score fusion produces incorrect rankings</li> <li> <strong>RAG-116:</strong> Metadata filter ignores date conditions</li> <li> <strong>RAG-117:</strong> Reranker returns raw scores instead of final documents</li> <li> <strong>RAG-118:</strong> Faithfulness evaluator always returns 0%</li> </ul> <h3> 2. Write Code in a Real Editor </h3> <p>The workspace has:</p> <ul> <li>A task description panel (the ticket)</li> <li>A Monaco code editor with starter code</li> <li>A terminal panel that shows results</li> </ul> <h3> 3. Automated Evaluation </h3> <p>When a user submits code:</p> <ol> <li>It runs in a sandbox (timeout + memory cap)</li> <li>Pytest suites check correctness</li> <li>An AI judge (Groq LLM) provides detailed feedback</li> </ol> <h3> 4. Verified Skill Profile </h3> <p>Users get a talent graph showing scores across 5 competencies:</p> <ul> <li>Chunking</li> <li>Fusion</li> <li>Filtering</li> <li>Reranking</li> <li>Evaluation</li> </ul> <h2> The Tech Stack </h2> <ul> <li> <strong>Frontend:</strong> Next.js 16.2.10, React 19, TypeScript, Tailwind CSS 4</li> <li> <strong>Editor:</strong> @monaco-editor/react</li> <li> <strong>Backend:</strong> Python 3.13, FastAPI</li> <li> <strong>Database:</strong> Supabase (PostgreSQL)</li> <li> <strong>Auth:</strong> Supabase Auth</li> <li> <strong>AI Judge:</strong> Groq API (llama-3.3-70b-versatile)</li> <li> <strong>Hosting:</strong> Vercel (frontend), Render (backend)</li> </ul> <h2> The Architecture </h2> <p>The backend pipeline is a plain Python function composition:</p> ai programming tutorial startup