The Rule: If an AI coding assistant (like Cursor or Copilot) can build your project in one prompt, hiring managers do not care about it. You must build systems with complex state, orchestration, and concurrency.
Project 1: The Distributed RAG Pipeline
A basic RAG pipeline is trivial. Build a system that can handle massive enterprise data.
- Requirements: Ingest 1,000 page PDFs. Chunk the text semantically. Spin up a Celery/Redis background worker to generate embeddings asynchronously so the main API doesn't block.
- The Hard Part: Handle document updates. If page 5 of the PDF changes, how do you update the VectorDB without re-embedding all 1,000 pages?
Project 2: The Agentic Cost Router
Companies are spending millions on API costs unnecessarily. Build a proxy gateway that saves them money.
- Requirements: Build an API gateway that intercepts all prompts going to OpenAI. Use a fast, local Small Language Model (like Llama-3 8B) to classify the difficulty of the prompt.
- The Hard Part: If the prompt is "easy" (e.g., summarize this paragraph), route it to a cheap API (Claude Haiku). If it is "hard" (e.g., write a Python script), route it to GPT-4o. Prove your gateway saves 60% on costs over 10,000 simulated requests.
Project 3: The Multi-User Voice SFU
Text is dead. Real-time voice is the next frontier.
- Requirements: Build a WebRTC Selective Forwarding Unit (SFU) where 3 humans and 1 AI agent are in the same voice channel. The AI agent must listen to the humans, transcribe via Whisper, and interject contextually using TTS (Text-to-Speech).
- The Hard Part: Managing latency and preventing the AI from interrupting humans mid-sentence.