AI Lab

What It Is

AI Lab is a specialized in-browser development environment engineered specifically for modern technical evaluations. Rather than asking candidates to invert binary trees in a blank text box, companies use AI Lab to evaluate how candidates actually work with AI coding assistants to solve realistic, multi-file software engineering tasks.

Candidates receive a problem statement, a full browser-local filesystem, a Monaco/VS Code-like editor, an AI chat panel, a terminal, and live execution previews.

---

What It Does

• Interactive AI Coding Agent: Candidates solve programming challenges in dialogue with an AI agent capable of reading workspace files, generating code, and explaining architectural choices.
• In-Browser Virtual Filesystem: Multi-file projects run entirely in the candidate's browser via IndexedDB persistence without lagging roundtrips.
• Isolated Code Execution: Integrates isolated sandboxes (via E2B) for running actual backend scripts, compilation checks, and unit tests safely.
• Live Previews & Visual Tests: Web projects render instantly in an integrated live preview pane.
• Evaluator Governance: Enforces assessment parameters such as time limits, token consumption budgets, interaction logs, and scoring criteria.

---

Architecture & My Contribution

I was solely responsible for architecting and building AI Lab from scratch at Concept Ninjas:

1. System & Web IDE Architecture:

   • Designed the entire client-side IDE layout, editor state, tab management, and split-pane viewports.
   • Built the browser-local virtual filesystem engine (later published as the idb-vfs NPM package).

2. Model Abstraction Layer:

   • Architected unified LLM connectors allowing pluggable integration with Gemini, OpenAI, and Claude without altering client application logic.
   • Implemented streaming token responses and conversational context pruning.

3. Secure Sandbox Orchestration:

   • Engineered secure execution pipelines connecting client code to remote E2B sandboxes for isolated code execution.

---

Skills Used & Developed

• System Design & Browser Architecture: IndexedDB state persistence, web workers, and Monaco editor integration.
• AI Agent Orchestration: Tool calling, context window optimization, and prompt guardrails.
• Full-Stack Performance: Next.js, Redis session caching, and sub-second terminal responses.

---

Key Learnings & Takeaways

• Scale in assessment platforms: The platform now powers over 500 live real-world assessment sessions every week with zero data corruption or lost candidate code.
• Candidate ergonomics matter: Candidates under testing pressure cannot tolerate editor quirks. Investing heavily in local VFS reliability directly translated into seamless candidate experiences.