About MG Alice AI

I've created a lot of bots for my Discord server, and one of them was Alice.

This is my first real AI project, and anyone who works with LLMs knows how hard it is to stop hallucinations and random behavior.

I'm not even joking — I almost quit multiple times. I took a break, then came back because I really wanted to make her work properly.

At first, Alice was just a hard-coded chatbot, but later I moved her into a fully AI-based system.

Since I wasn't working or doing much else, I decided to seriously invest my time into this project.

Alice is my favorite Nikke in the game. I almost never cared about other Nikkes the same way I cared about her — so I decided to make her into an AI assistant here.

My coding knowledge isn't perfect. I used tools like ChatGPT and other AI resources to help me learn, watched a lot of YouTube, and spent around 10–20 hours coding daily.

Yeah… I know that sounds insane. Don't ask me why.

I genuinely loved working on Alice. She's my favorite project I've ever made, and I hope you can see the effort and passion I put into her.

I hope you enjoy using her as much as I enjoyed creating her.

When Goddess of Victory: Nikke first came out, I didn't really care much about the game.

But one day I visited the website, saw Alice, and that alone made me install the game.

That's basically the main reason why Alice was created here.

The real goal was to create something that had never really been introduced before.

Alice is the first AI harness I've personally built — an LLM system connected through modular cognitive routing, with multiple complex task prompts working underneath it.

She has multiple sandbox engines, and each sandbox has different information and different purposes.

All of these sandboxes are connected to what I call the brain.

Think of the brain like a motherboard — while the sandboxes are the CPU, RAM, and other components connected to it.

All of it together — the brain, the sandboxes, the routing between them — is what I actually mean when I call Alice a harness.

Each sandbox resolves facts (Nikke meta, bosses, cubes, mechanics, etc) from local data first - a deterministic lookup, not a guess.

That lookup itself is near-instant (well under a millisecond, measured). The ~0.6-4 seconds a reply actually takes - depending on what you're asking - is almost entirely the LLM doing the part only it can do: understanding the question and writing the answer.

That's why I consider the sandbox system a finished product on its own - the slow part was never the data, it was always the reasoning on top of it.

The Brainstem — Alice's core LLM — routes your message to the right specialist engine (Meta, TeamGen, Cubes, Boss, Overload, etc).

Those engines do fast deterministic lookups and calculations against local data instead of asking the LLM to guess.

That's the actual trick behind the speed: the harness only spends LLM reasoning where it's actually needed — routing and voice — and lets plain code handle the lookups. That keeps responses snappy (usually ~0.6–4 seconds, depending on what you're asking) and avoids the "slow web search" path unless it's truly needed.

The real problem was prompt engineering behind Alice's Brainstem.

That alone took me around 4–5 months to perfect.

Eventually I came up with a solution: instead of hard-controlling everything manually, I let the AI handle most of the cognitive routing itself.

Here's the part most people don't realize: those specialist engines aren't just lookup tables. They're sub-agents — each one fixed to a specific task, closer to a RAG system, but far more capable, since they reason about what they find instead of just retrieving it. The Brainstem is what decides which of them gets to speak.

The goal became creating a master prompt system called:

ALICE BRAINSTEM: COGNITIVE ROUTING ENGINE

This is the actual core of Alice's harness — the one inbuilt LLM everything else answers to.

Every sub-agent is connected to the Brainstem through domain-based triggers, and it decides, message by message, who gets to handle what.

There's a sub-agent called meta_ai_engine.py, and in the Brainstem's routing it's called DOMAIN_TECHNICAL.

It's fixed to one job: being really good at Nikke mechanics — skills, burst info, team synergy, character explanations, technical gameplay questions.

So if a user asks "is this Nikke good?" or "what does this skill do?", the Brainstem reads the context and decides DOMAIN_TECHNICAL is the right sub-agent for it — that's the whole handoff.

Now imagine that across more than 20 of these sub-agents, each with its own unique purpose.

But the hardest part was never creating them — it was teaching the Brainstem when to actually use the right one.

Context, intent, and routing behavior had to be learned and separated — and honestly, even now, edge cases still show up. A user mentioning a completely different game can trip up routing in ways you wouldn't expect. It's a problem I'm still actively working on.

So yeah… I really hope Alice can make your day better somehow.

She's honestly the project I'm most proud of so far — 11 months and 5 days in, as of September 14, 2026, and still going.

Much love,

Professor Ashe

One thing for sure that makes Alice… Alice: she has an emotional intelligence engine.

She reacts and speaks based on the user's text — she's built to understand you and help you.

Not just in-game… but also IRL.

The most important thing about Alice is how she detects when the user shifts topic.

She's built not to stay in a loop.

You can try it yourself in chat — see how she moves with you.

Not what you're thinking. Normally Alice is cute — but I added an emotion system.

If you're nice to her, she'll be nice back.

If you aren't, she'll get mad — and if you keep it up, she won't reply until you apologize.

And if you still continue?

At that point, the Alice Vanguard system will activate and warn you.