AI Chat Application
You're the engineer of record for a startup's first AI feature: a website chat assistant. Not a demo — something real users will actually use: it answers fast, doesn't forget the conversation, and won't surprise anyone with a bill.
What you'll build
A streaming, multi-turn chat application with a designed persona, robust error handling, context management, and a per-conversation cost budget.
Why it matters
The atomic unit of applied AI work. Every AI-touching role assumes you can do this well — most candidates do it badly (no streaming, no truncation, no cost thinking).
Skills you'll gain
Prerequisites
Path gates M0–M3, a dev environment, and an LLM API key.
Requirements
- RQ-1: multi-turn coherence over ≥10 turns.
- RQ-2: first token < 2s perceived via streaming.
- RQ-3: graceful degradation on API errors — the user never sees a stack trace.
- RQ-4: conversation cost tracked and capped.
- RQ-5: persona holds under pressure.
Architecture
Product-level: client ↔ conversation manager (state, truncation policy) ↔ LLM API adapter (streaming, retries, provider-swappable) ↔ cost meter. The explanation teaches why a conversation is an engineered object, not which framework to use.