Case study
An AI video platform, from an empty AWS account to production
Semantic search across thousands of TV-length videos: upload, transcode, enrich with AI models and index, built and run end to end as the founding engineer.
Project descriptions are intentionally high-level to respect employer confidentiality.
Results
- production videos searchable by meaning
- 3,500+
- lower cost per reprocessing run
- 30x
- lower transcoding cost with optimized encoding profiles
- 93%
- automated tests, 80%+ coverage
- 6,600+
The problem
- Customers upload very large files (up to 100 GB) that must never pass through our servers.
- Every video has to become searchable by meaning, which means several AI models per segment, each slow and quota-limited.
- Models and prompts change often, so reprocessing the library has to be cheap.
Decisions and trade-offs
Orchestrate work that mostly waits
Most of the pipeline time is spent waiting on AI models. A managed workflow engine gives per-stage retries and a visible state for every video, without paying for workers that sit idle.
Idempotent by design
Every stage can be re-run safely: results are written so that a retry replaces earlier output instead of duplicating it, and a video only becomes searchable once indexing succeeds.
Send each stage only the data it needs
Passing whole source files between stages caused timeouts on very large videos. Giving each stage only the data it actually uses removed them and cut transfer and processing overhead.
Treat model quotas as a design constraint
Model calls are rate-limited, so concurrency is kept below the quota and excess work waits in a queue. Backpressure instead of retries keeps a burst of uploads from turning into a retry storm.
Cache every stage
Stage-level caching means changing one model or prompt only recomputes the stages it affects. That is what made each reprocessing run 30x cheaper.