We've built this kind of pipeline before.
Most of what runs underneath an AI video editor is the same plumbing we've spent years writing for backend AI work. Ingest media at scale. Run inference. Turn messy model output into something a person can review and trust. The shape changes when you point those pipelines at video, but the bones are the same.
Production AI also taught us where these systems fail. The first failure is removing the human review step and shipping whatever the model decides. The second is hiding the review step in a UI that wastes the user's time. Both failures are visible right now in the AI video editor market, which is why Sapari exists — the AI proposes, the timeline shows you cards, you decide.
Built with creators, not at them.
We started Sapari by sitting with a small set of creators — YouTubers, short-form posters, course makers, coaches — and watching how they actually edit. We didn't run surveys. We watched what they did at midnight when something needed to ship by morning.
Every feature in Sapari traces back to something we saw. Someone scrubbing through 4 minutes of silence to find the next sentence — that's why the pacing slider goes from Off to Hyper. Someone re-cutting the same false start three times — that's why false-start detection ships as colored cards on the timeline, not as a black-box auto-trim. Someone exporting the same edit four times for four platforms — that's why one project becomes 16:9, 9:16, and 1:1 from the same timeline.
We're not trying to remove you from the edit. We're trying to remove the parts of the edit you didn't want to be doing in the first place.
Three years of open source.
The libraries below are the public surface — the part you can `pip install` and read on GitHub at benavlabs. There's more behind them: AI agents in production, infrastructure templates, internal tooling we built shipping enterprise AI projects.
FastroAI is split in two. The package above is open source — agentic orchestration on top of PydanticAI. The closed-source half, FastroAI Template, is a full-stack starter for production AI apps and was our first paid product. Both live at fastro.ai; documentation for both is at docs.fastro.ai.
Sapari sits in a different place. FastroAI Template was for developers; Sapari is the first product we've built for an audience that isn't. The standards carry over either way — ship something we'd use ourselves, document it honestly, don't pretend the model is doing more than it is.
Who's building this.
Four of us, for now.
Igor Benav
Author of FastCRUD and FastroAI. Builds production AI pipelines for a living — billions of records ingested, BI dashboards serving 100+ daily users, the orchestration layer Sapari runs on. Teaches Python and MVP development on the side; plays guitar and runs D&D campaigns when he's not at a keyboard.
Vic
Runs product and operations at Sapari. Spent years at Benchpark managing art and creative deliveries for Disney, Nubank, Zombie Studios, and Fini — knows what shipping creative work on a real deadline actually looks like. Digital games designer by training, top-of-class honors at IESB.
Gabriel Silva
15+ years shipping production software. Built the AI chatbot engine inside Caixa Tem (Brazil's largest digital bank), ANVISA's pharmaceutical registration system, banking automation for Banco da Amazônia, and risk-calculation APIs serving 30K+ monthly requests. Production AI where "the model got it wrong" has consequences.
Lucas Rodrigues
Owns QA at Sapari and maintains the open-source work at Benav Labs — issues, PRs, releases on FastCRUD, CRUDAdmin, and the rest. Python and FastAPI background, top-4 finalist at Campus Mobile (Brazil's largest telecom entrepreneurship competition).