Maitri Vaghasiya

What if Apple designed YouTube?

got asked this in a YC company interview system design round recently "how do you keep an AI system aware of a project after 100s or 1000s of edits, without sending the entire history every time, and users say vague stuff like “make that bigger”" most people try to solve this with bigger context windows or by stuffing more prompts in, but that breaks as noise increases and tokens get expensive what i did instead was stop treating prompts as plain text and treat every interaction as a structured action over a canonical state so instead of the model trying to remember everything, the system maintains a source of truth state that always reflects the latest version flow was simple: parse intent, resolve references like “that” using current state, run a confidence check, then execute and update state the reference resolution step handles most of the ambiguity by mapping vague inputs to concrete entities instead of leaving it to the model to guess i also added a confidence layer so if the system is unsure, it asks the user instead of making silent wrong edits every action is stored as an event, not to replay everything, but to retrieve only the relevant history when needed for context building, i send the current structured state plus top k relevant past events using semantic retrieval, which keeps tokens low and signal clean main shift was this: don’t make the model remember everything, make the system decide what matters and feed only that

Day 1: 🌱 Day 100: 🌷 Day 365: 💐