JioCreate
A multi-agent platform that turns a creator's brief into a finished short video. I led its architecture: 20 specialised agents across a seven-stage pipeline, automated quality gates that hold every script to 8.1 out of 10, and an Audio Signature System that keeps the generated voice the same from one 8-second scene to the next. Production went from weeks to hours.
- My role
- AI product manager: architecture, quality system and model strategy
- Team
- Cross-functional AI product team of 14
- Users
- Content creators producing short videos at scale
- Stack
- Google ADK, Gemini 2.5 Pro and Flash, Veo 3.0, Imagen 4.0
The Problem
A short video needs research, a script, a shot plan, images, video, a consistent voice and captions. Done by hand that takes a team weeks. Done in one large model call, quality is hard to control and the narrator's voice changes from scene to scene. The job was to automate the whole chain and still publish only work that clears a quality bar.
How It Works
A workflow orchestrator moves each brief through seven stages, handing work to specialist agent teams for the creator profile, script, direction and video. Typical time per stage:
- Profile~2 s
- Research5–10 s
- Script15–30 sQuality gate · up to 2 rewrites · creator approves
- Production notes20–40 sQuality gate · up to 3 rewrites
- Images30–60 s
- Video2–5 min
- Captions
Outlined stages have a quality gate: critic agents score the work and refiner agents rewrite it until it clears the bar or the rewrite limit.
- Quality gates. Critic agents score each script and production plan on six dimensions: engagement, clarity, persona alignment, research accuracy, structure and trending relevance. Anything under 8.1 out of 10 goes back to a refiner agent.
- Model mix. 6 agents run on Gemini 2.5 Pro for the judgement-heavy work, such as writing scripts and planning production; 14 run on Gemini 2.5 Flash for feedback and refinement.
- Voice consistency. Veo generates video in 8-second scenes, so the narrator's voice can drift from scene to scene. The Audio Signature System checks every scene against 10 voice parameters with set tolerances.
- State across stages. ADK callbacks (before and after each agent and tool) carry parameters and the creator's feedback forward, so a change approved at the script stage shapes the production plan without losing context.
Decisions and Trade-offs
- Specialist agents over one big call. Splitting the work let each agent do one job well and let us put the expensive model only where it changed the result. The cost is orchestration complexity, which the workflow orchestrator and callbacks absorb.
- Bound the rewrites. Early versions refined until a script cleared the bar, which ran up cost and gave little after the third or fourth pass. Capping rewrites at 2 for scripts and 3 for production notes cut average generation time by 40% while holding the 8.1 bar.
- Pro where judgement matters, Flash everywhere else. 6 of the 20 agents run on the larger model. Quality is set by the script and the shot plan, so that is where the spend goes.
- Humans approve at the points that matter. The creator approves the script before production planning starts. Automated gates catch weak work; people make the calls that are about taste and intent.
- Measure the voice, don't eyeball it. Voice consistency is checked as parameters with tolerances rather than by listening, so it can run on every scene.