Talecut
Co-built Talecut, turning long-form horizontal video into short vertical clips; reached 400+ active users in three weeks.
Co-built and publicly launched Talecut in a team of two during Bitmovin’s 10-week AI Incubator, building on the company’s AI Scene Analysis technology. The product turns long-form horizontal video into short vertical clips.
The modular pipeline combines LLMs, vision and speech models to select moments, track subjects for the vertical crop, and generate transcriptions and subtitles. A Python backend and React frontend connect the processing pipeline to a public-facing product at talecut.com.
We shared the engineering and launch work, from the Python backend and React frontend to computer vision and infrastructure on GCP, Kubernetes and Terraform. Redis workers distribute processing across CPU and GPU resources, with shared leasing, caching, cancellation guarantees and replayable stages. Resolution testing improved throughput by 2.6× with no measured quality loss.
Highlights
- Reached 400+ active users in three weeks after launch.
- Improved throughput by 2.6× through resolution testing, with no measured quality loss.
- Designed a 17-stage pipeline with shared leasing, caching, cancellation guarantees and replayable stages.
What I did
- Co-built the Python backend, React frontend and computer vision stack, including shot detection, video segmentation and camera framing.
- Operated Kubernetes infrastructure on GCE L4 GPUs, with DGX Spark for local development.
- Resolved capacity shortages, GPU driver failures and read-only hosts.
- Introduced generated cross-language contracts, CI/CD, testing and run telemetry; identified scheduling as the main performance bottleneck.
- Shared product naming, branding, website and SEO work, ran a paid acquisition campaign, and demonstrated Talecut to prospective clients.
Stack
- Python
- React
- Redis
- Kubernetes
- GCP
- Terraform
- NVIDIA L4
- PostgreSQL
- Object storage
- CI/CD