PerfLens: interactive LLM inference performance sandbox

Help new grads and career switchers understand prefill/decode, KV cache, and other internals through hands-on simulation

Problem

Interviewers find many fresh graduates can recite QKV theory but fail when asked what KV cache stores, why long contexts are slow, or why token billing differs. School teaching and interview prep rely on AI-generated surface answers, lacking hands-on experimentation.

Solution

A browser-based interactive sandbox that visualizes LLM inference: shows computational differences between prefill and decode phases, dynamically displays KV cache growth, lets users adjust context length and batch size to observe latency and memory changes in real time, with explanations and interview question cards.

Users

New grads preparing for LLM-related job interviews, career-switching developers, and engineers wanting deeper understanding of inference

MVP

Simplified simulator based on a small open-source model; visualize prefill/decode and KV cache; offer several preset scenarios with corresponding interview questions; support note export.

Revenue

Free basic scenarios; subscription for more model configurations, interview question bank, and note export; or one-time purchase.

Why now

LLM job interviews are highly competitive, yet teaching tools for low-level principles are scarce. Browser WASM and visualization libraries are mature, enabling indie developers to build at low cost.