AI Demo Video Builder
A multi-model agent pipeline that turns a silent screen recording into a narrated, captioned product demo. Four models across eight phases - one perceives the video, one directs the edit, two render it - with deterministic code holding the seams.
8-phase agent pipeline · 4 models orchestrated · 410 tests passing
the problem
Turning a raw screen capture into a product demo means watching it back, deciding what matters, cutting it, writing narration that matches what is on screen, and rendering voice and avatar in sync. Handing all of that to one LLM produces confident nonsense: hallucinated features, drifting timestamps, and ffmpeg commands that fail halfway through a render.
what i built
I split the system on a hard line - AI makes editorial decisions, deterministic code executes them. Interfaze perceives the video and returns structured observations; a reconciliation phase with zero model calls merges coarse and refined timelines and grades every claim as observed, strongly inferred, or product context, rejecting anything unsupported; Gemini 3.1 Pro then directs across six reasoning stages and emits a validated EditPlan that passes 32 checks before a single frame is rendered. ElevenLabs and HeyGen render, ffmpeg composites, and a verification phase runs deterministic checks with bounded revision loops.
why it's interesting
This is the project that best shows how I think about AI systems. Most of the hard problems were not prompting - they were failure modes you only find by shipping. Per-line narration produced 32.3 Hz of pitch drift between takes, so narration batches into a single take. Per-block avatar rendering made the presenter vanish during planned pauses, so it renders once over the whole narration bed. And no phase is allowed to discover an error mid-ffmpeg. Models are treated as unreliable components inside a system engineered to contain them.
built with
python, fastapi, gemini 3.1 pro, interfaze, elevenlabs, heygen, pydantic, ffmpeg, react, typescript