Patricija Žemaitytė argues that useful AI systems need infrastructure for fresh external information alongside their trained models. Three engineering stories examine how an initial data request can expand into a broader pipeline, how interactive retrieval changes latency requirements and why testing becomes difficult as traffic grows. The examples are the speaker's account of work at Oxylabs, rather than independently verified performance benchmarks.
Patricija Žemaitytė describes a video-data request that grew beyond downloading files to include transcripts, subtitles, language search, metadata and channel information. A misunderstanding between transcripts and subtitles illustrates why successful infrastructure depends on clarifying the data a downstream workflow actually needs. The lesson is that multimodal training inputs require coordinated collection, storage and delivery rather than a single downloader.
Patricija Žemaitytė contrasts broad search-result scraping with a narrower retrieval service designed for AI workflows. A reported sub-second prototype failed during a client test after access was blocked, forcing another iteration that relied more heavily on browsers. The subsequent work examined parsing, sessions and other infrastructure components to reduce cumulative delay. The case distinguishes a promising development result from a system that continues working under real operating conditions.
Patricija Žemaitytė explains that a scraping request can involve routing, rendering, browser execution, parsing, retries, normalization and delivery, so request counts alone conceal substantial work. During scaling, realistic client traffic was harder to reproduce than synthetic load, and logs and metrics became part of the system's own workload. The talk emphasizes gradual scaling, useful observability and ongoing adaptation as requirements change.
Watch on YouTube
