# AI (“Anak Intern”)

> How We “Onboarded” an LLM to Become a Smart Co-Worker. Treating an LLM like a new intern: hiring the right model, briefing it, and giving it memory with RAG — to build a Gemini Live API call assistant that feels human.

Canonical page: https://laam.my.id/talks/ai-anak-intern
Last updated: 2026-09-29

- **Type:** Talk
- **Event:** AInnovate 2025 · GDG on Campus BINUS University Online
- **Date:** 2025-11-01
- **Location:** BINUS University Anggrek Campus, Jakarta
- **Audience:** Students & developers
- **Language:** Bahasa Indonesia & English
- **Topics:** Applied AI, Gemini Live API, RAG, Vector Search, Function Calling, Flutter
- **Slides:** [view the deck](https://docs.google.com/presentation/d/1cTmHErqjlRCb99RNT6KQYimoSgl576Qjub9jXThnn14/edit?usp=sharing)
- **Source:** [github.com/GDGoC-BiOn/ainnovate-llm](https://github.com/GDGoC-BiOn/ainnovate-llm)

An LLM behaves a lot like a new intern: capable, but it needs the right hire, a clear brief, and context before it becomes useful. This AInnovate AI-stage session walks through that onboarding step by step.

Step one is "hiring" — choosing a model size. Step two is the briefing: temperature and a system prompt that turns the model into an empathetic mood companion. Then comes the real problem — a small model has a small context window, so it keeps asking "sorry, who are you again?"

The fix is RAG: grounding the assistant in WHO guidelines through vector search, and — because Gemini Live API works on audio streams — wiring retrieval in through function calling. The result, demoed live in a Flutter sleep-call app, is not a chatbot but a co-worker.