princeton CS. building AI tools. i post what i use.

last updated June 9th, 2026

<aside>

13 builds ranked by how hard they make a recruiter stop scrolling. each one ships with the exact Claude Code prompt to build it. pick 3, not 13.

</aside>

<aside> ⚠️

the one rule that makes or breaks all 13. u have to be able to DEFEND every design decision in the interview. vibecode the boilerplate, hand-verify the core logic, write the README urself.

</aside>

<aside> 🧰

every prompt below is a Claude Code mega-prompt. before u run one, add ur own constraints, the language u want, the GPU u actually have, and ur deadline.

</aside>

1. mini llm inference engine (mini-vLLM)

the from-scratch LLM serving engine. it loads a small open model (Qwen2.5-0.5B or Llama-3.2-1B) and serves it with the two tricks that made vLLM famous, paged KV-cache attention and continuous batching, behind an OpenAI-compatible /v1/completions endpoint.

this is the systems side of LLM serving, which is exactly what AI infra teams hire for. almost nobody early-career has built it. thats the point.

what it has

• model loader, HF safetensors into ur own PyTorch forward pass
• a KV cache manager with fixed-size pages and block tables (PagedAttention)
• a continuous batching scheduler that admits requests mid-flight and evicts finished ones
• sampling (temperature, top-p) and streaming over SSE
• a benchmark harness, throughput in tok/s and p50/p99 latency vs naive model.generate()

<aside> 💡

how to stand out. publish the benchmark table, naive HF generate vs ur engine at 1, 8, and 32 concurrent requests. then write the post i built vLLM from scratch to understand PagedAttention and point ur AI-lab apps straight at it.

</aside>

• 📋 the claude code prompt (click to copy)

2. train a small llm from scratch (nanoGPT-style)