Sanford Holsapple

A creative system built at Netflix that programmatically compiles personalized, multi-title promotional video reels ("Dynamic Sizzles"), replacing a fully manual, non-personalized production process. Co-invented and patented as "Methods and Systems for Providing Dynamically Composed Personalized Media Assets" (US Patent 12,177,542).

• Designed the core logic — architected how footage needed to be structured so any title could be eligible for any position within a sizzle, scalable across titles, genres, music tracks, and future use cases
• Defined the clip hierarchy (parent clips with duration variants), the cadence system, and the hardcoded assembly rules the system required
• Evolved the system into a production creative-direction tool that let editors and strategists configure cadence and single- vs. multi-title experiences themselves, with no engineering support needed

The system depended on a large volume of metadata staying accurate from editorial through encoding to final assembly, and manual entry was producing errors. I identified the metadata requirements and specified how a Premiere Pro plugin should operate to solve it — scraping metadata directly from source clips and embedding it as Premiere markers, so it could flow into the EDL automatically with no manual re-entry. An engineer built the plugin to that spec, removing the error source entirely.

Invest more in AI-assisted creative input. The requirements for each moment in a sizzle are well-defined enough to codify, so a model could suggest the strongest candidate moments rather than relying on fully manual curation.

An end-to-end workflow at Netflix that used AI to convert tens of thousands of existing horizontal video assets into vertical format with minimal human intervention, replacing a fully manual, agency-led adaptation process. Co-invented and patented as "Automated Video Cropping" (US Patent 11,477,533).

• Architected the overall solution and core logic
• Evaluated manual and automated approaches, assessed how different deal structures would affect budget and timeline, and audited output quality from internal and external models to land on an AI-first, minimal-review approach
• Diagnosed why algorithm output was inconsistent across genres and shot types, and proposed the fixes that raised acceptance rate from 31% to over 70%
• Shaped the editing tooling editors used to adjust AI output, and partnered with Data Science & Engineering to confirm the results didn't harm member experience
• Scaled the resulting workflow to 17,000+ titles across four global regions, reaching ~67% automated vertical video coverage

This work started with no product test attached to it, so there was no natural anchor for investment. I had to make the internal case that vertical video was worth building for before there was a committed use case — sequencing asks to engineering so the initiative kept moving without pulling resources from higher-priority work, and re-adjusting the plan each time a planned test got postponed. When a strong test opportunity came, the groundwork meant we could move into it quickly.

Push earlier for direct delivery from the internal editing tool to the CDN. Without it, editors had to manually download and deliver finished vertical videos, so the last mile of the pipeline stayed manual even after everything upstream was automated.

An internal Netflix workflow using GFX templates to enable dynamic, automated localization of burned-in text within video — so an original-language graphic could be automatically replicated across languages, encoded, and delivered without manual redesign per language. What used to require 32 separate manual video deliverables per title now requires one.

• Formulated and pitched the concept, building on existing internal technologies and pipelines, and set the vision, use cases, and projected impact that got the project greenlit
• Ran scheduling and milestones throughout
• Served as creative director for the GFX template system and its web editor — defining which parameters needed to be adjustable (stock motion, motion augmentation, font sets) and what tooling features were necessary for the system to work across the full range of titles

Once live, editors reused the same design elements far more often than expected, creating pressure to constantly refresh the templates. Rather than treat it as a design problem, I addressed it structurally: adding the ability to upload custom GFX backgrounds (static or video), and enabling alpha-channel rendering so the background could be embedded directly in the AV file and only the overlaid text needed to be replaced — reducing the system's dependency on a limited template library.

Split the work into three sequential phases instead of running them concurrently: first assess feasibility against existing tech/pipelines, then build and stress-test the template system in isolation (e.g. handing it to motion designers in After Effects to find gaps), and only then build the web app and end-to-end pipeline once the template was locked. It would have taken longer, but avoided rework on both the templates and the web app.