SAGEN AI

Overview

SAGEN AI is an AI SaaS platform. Users pick an avatar from a selection and talk to it \u2014 ask questions, get responses back in the avatar\u2019s voice, with synced lip movement and body language. Think of it as a conversational AI with a face.

Problem

AI agents in 2023 were mostly a chat window. You typed, it typed back. Functional but flat. No one felt connected to the experience. SAGEN wanted to change that \u2014 give the AI a presence, a personality, a face that users could actually talk to. The challenge was building a full animation pipeline that could take a static 2D avatar and make it speak, emote, and move naturally, all synced to real-time audio.

Process

The Live2D Pipeline

The biggest technical challenge was getting Live2D assets into Unity. The standard Live2D workflow doesn\u2019t talk to Unity out of the box. I had to build a custom pipeline: export from Live2D Cubism, convert the rig data, write Unity shaders that could interpret the mesh deformations, and sync everything to the audio timeline.

1. Slice the avatar \u2014 Started from a static render PNG. Cut it into separate parts \u2014 eyes, mouth, brows, torso, arms \u2014 each becomes its own layer in Live2D.
2. Rig in Live2D Cubism \u2014 Set up deformers, parameters, and blend shapes. Every part needed to move independently so the avatar could gesture and emote.
3. Animate visemes \u2014 Mapped mouth shapes to phonemes so the lip-sync works in real time. Each viseme had to match the audio waveform frame by frame.
4. Export to Unity \u2014 Wrote a shader-based renderer that could handle Live2D\u2019s mesh structure, texture atlases, and animation parameters. This was the hardest part.

QA and Staging

I handled QA for all of this in Unity and on staging. Testing viseme sync frame by frame, checking pose transitions for clipping, verifying audio latency across devices. Every avatar had to look good in every language\u2019s phoneme set.

Web and App Design

Outside the animation pipeline, I designed the SAGEN website and the mobile app. The main piece was the user dashboard and the avatar selection flow \u2014 users pick an avatar, customize a few settings, and start talking. The interface had to stay out of the way so the avatar experience stayed front and center.

Solution

The final product had:

• Avatars that speak with synced lip movement and natural body language
• A Live2D-to-Unity pipeline that turns static 2D art into real-time animated characters
• Viseme mapping for accurate real-time lip-sync in multiple languages
• A clean dashboard and avatar selection flow across iOS, Android, and Web

Results

Users felt connected to the AI. They preferred talking to an avatar over typing \u2014 it wasn\u2019t a gimmick. In the early days, the avatars and their animated presence were the primary differentiator in a market full of text-only AI assistants.

Learnings

The Live2D-Unity bridge was harder than expected. The tooling wasn\u2019t designed for this. If I did it again, I\u2019d push for a deeper investment in the animation pipeline earlier \u2014 the viseme system alone took multiple iterations to get right. But the result was worth it: avatars that actually feel present in the conversation.