AI Character Live Runtime
Building AI characters that can perform live without getting tired.
A live runtime for AI Vtuber-style characters that can stream, publish videos and shorts, remember viewers, follow scripts, and run fully automated shows.
4
active AI Vtubers
3000+
watch hours
24h
stream potential
Auto
live performance
The Problem
Human Vtuber performance is constrained by fatigue, scheduling, memory, and consistency. The challenge was making AI characters feel alive while also being stable enough for real livestream operation. The hardest part was never reading a script out loud — it was generating the stream agenda. Early agendas were too formulaic and too obviously machine-written to sound like the character, and it took many rounds to reach the current flow: start from the character, find a topic, generate the agenda, deliver the script, read the chat, and go off-script when the moment calls for it.
Role: Project Lead · Persona Designer · Runtime Debugger · Live Flow Owner · Tooling Integrator
These 4 characters are real and currently in active operation — the cards are a recreation, but the roster and the numbers are not.
Illustrative — not a real screenshot
AI Streamer 01 · live now
412 watching · started 18m ago
AI Streamer 02 · live now
198 watching · started 42m ago
AI Streamer 03 · live now
356 watching · started 7m ago
AI Streamer 04 · live now
287 watching · started 53m ago
3000+
watch hours — accumulated from real viewers, not a lab demo.
Before
- Characters could exist as concepts, but live operation required many fragile manual pieces.
- Persona, tooling, runtime behavior, and livestream flow were not yet unified into a stable performance system.
- Long-running shows needed better consistency than a human operator could maintain manually.
- Agenda generation started out formulaic and obviously machine-written — not anything the character would actually say.
After
- Four AI Vtuber characters are in active operation — Dearu (小鹿獸), NaisKuri (奶糖狐), AkuCat (惡貓社長), and Amber (琥珀喵). All are company-owned, designed under my lead and built together with colleagues.
- The system has accumulated more than 3000 watch hours from real viewers, cumulative across livestreams, videos, and shorts.
- AI characters can run stable, fully automated livestream performances: the agenda is generated in the character voice, the character reacts to chat, and it goes off-script when the moment calls for it.
- A collaboration with Gamania on their virtual talent 宮祈緣: before the fully-AI version went live we ran an A/B test of a human performer against the AI, and both sides agreed the AI ran it better.
one stable runtime
Contributions
What I actually did.
Led project management and execution across persona, runtime, and live workflow.
Designed character personas and operating behavior — values, preferences and behavioural rules, down to the precise boundary of what the character should refuse to do.
Debugged runtime issues and livestream flow stability, and redesigned the agenda generation flow to get rid of the machine-written tone.
Defined character settings, tool connections, and automated show behavior.
Coordinated the system toward demo-ready and production-like livestream performance.
AI collaboration governance
Keeping AI characters live without incident isn't about writing a good script — it's about turning "which AI behaviors can't be left unchecked" into architecture and rules. This page is the governance I paid for in hard lessons.
Governance cases
Open for first-hand evidence
The AI's ingrained pattern
Early vibe coding: the character's stateless persona data and stateful runtime sessions all lived in one Node.js process's inline memory. An hour into a stream, as concurrency rose, context bloat caused an OOM and the character froze — amnesiac — live on air.
The governance mechanism I built
Broke decisively with vibe coding, separating identity, writing style, and the memory layer — stateless persona decoupled from stateful runtime — and built an observable runtime debugging path.
The collaboration value it unlocked
The lesson that a hollow prototype collapses at first contact with production forced my architecture thinking; that failure is the real starting point for the whole runtime and workflow I designed afterward.
The AI's ingrained pattern
I assumed generating a readable script was as simple as writing copy, but the AI-generated stream agenda had too heavy an AI footprint — too formulaic, no character voice, no flexibility. It produced "averagely correct" but soulless content.
The governance mechanism I built
Beyond iterating with front-line operators, I redesigned the whole generation flow so the agenda starts entirely from the character: pick topics, generate the agenda, deliver the script, read the chat, break from the script when it fits — every step bound to the persona.
The collaboration value it unlocked
Upgraded "can generate" into "a stream this character would actually run"; the flow now outputs performances with humanity and character consistency, not generic AI scripts.
The AI's ingrained pattern
AI defaults toward being all-obliging, answering anything — but that dilutes a persona. A character who can do and knows everything feels less like a real person.
The governance mechanism I built
Built each character from scratch with a fully-formed identity: values, preferences, behavioral norms — explicitly drawing the line on what the character does NOT do, extracting the human-felt "AI-ness" to suppress it deliberately.
The collaboration value it unlocked
A rarely-noticed governance dimension — making a character credible through what it refrains from, rather than exposing itself by being omnicapable.
Outcomes
Validated with real viewers
Not a lab demo: at the debut stream for Dearu, viewers who wandered in watched to the end without realising it was AI.
Always-on performance
Unlike human performers, AI characters can stream continuously, follow scripts precisely, and remember viewer context.
Commercial collaboration
The 宮祈緣 A/B test with Gamania concluded that the AI version outperformed the human-played one. The technology was licensed out; how the character is run from there is their call.