{"slug": "ai-s-streaming-stack-image-interpretation", "title": "AI's Streaming Stack: Image Interpretation", "summary": "OpenOrigins, a company that calls itself 'the verification layer for the internet,' uses AI to authenticate historical media content, with its Anchor product designed to prove the provenance of archived files and detect tampering. Founder and CEO Manny Ahmed said the company's biggest client is ITN, a U.K. video production house, and that large tech companies are seeking authenticated content for training AI models because they can no longer trust internet content to be AI-free.", "body_md": "# \n                    AI's Streaming Stack: Image Interpretation\n\nAI is a funny thing in that everyone likes to claim they use it in their products, but often during conversations with vendors, I find myself asking, “Is this AI or not?” (While they rarely say it’s not AI, sometimes its implementation seems marginal at best.) \n\nAnother issue that comes up is that not everyone wants to talk about what exactly they are using AI to accomplish, or they prefer to talk about future AI features rather than the capabilities of their currently available products or services. There’s not much point in trying to write anything worthwhile about vaporware or features that might sound great but can’t be tested or aren’t in use or available to users.\n\nThe companies covered here have substantially leveraged AI or incorporated it into their currently shipping tech. The representatives I spoke to discuss some very interesting approaches for what essentially is image interpretation, including detecting if content has been tampered with using AI, local AI processing, and [automated clip generation](https://www.streamingmedia.com/Articles/ReadArticle.aspx?ArticleID=172814) such as [vertical formatting](https://www.streamingmedia.com/Articles/ReadArticle.aspx?ArticleID=175220).\n\n## **OpenOrigins**\n\n[OpenOrigins](https://openorigins.com) is in the business of proving the source of imagery and data. It calls itself “the verification layer for the internet.” One of its services is authenticating historical content. [OpenOrigins founder and CEO Manny Ahmed](https://www.linkedin.com/in/humannyty/) offers this example: “How do we prove that that is the real version of Barack Obama’s speech from 2008 and not something that’s been generated today?” The company’s product, [Anchor](https://openorigins.com/products/archive-anchoring), is designed to answer questions like this. It is built to authenticate historical content stored in large media archives.\n\nAhmed says the way Anchor works is “quite simple. We do a security audit of the archive, then we establish trust in the timestamps associated with each media item. So, if a media file claims to be from 2008, we want to have three different datapoints that can give us confidence that it has not been manipulated since then. Once we’ve done that analysis, we take a visual summary of each media item—which we call an ‘origin hash’—and a proof of each media item, and we put that on the blockchain.”\n\nWhat are these datapoints? “Usually, an archive would be stored in two different tiers of storage—cold storage and hot storage—and there’s usually a bunch of services that hook into an archive,” Ahmed says. “So, you’ll have some sort of a logging or media compression service. You basically go through the logs for all of those and try to find the earliest references to each of those files.” This entails “running surveys on the client’s infrastructure,” he continues, rather than exporting “the content out. We don’t want a copy of the content.”\n\n*OpenOrigins’ Anchor matching using visual similarity (copyright OpenOrigins;* *used by permission)*\n\nAhmed says OpenOrigins’ “biggest client so far” is ITN, a large video production house in the U.K. “They produce videos for ITV, and almost all of their content going back 60 years is stored in an on-prem data center, but they use external services like a logging service and a media compression service, which gave us those datapoints,” he explains. “Say you find a random clip on Twitter that originated from your archive. We can then do a reverse-image lookup to prove that it came from your archive and also prove which parts of the image have been manipulated.”\n\nOpenOrigins is also helping provide authenticated content for training models. “A lot of the large tech companies that are training models say they can no longer get content on the internet that they can be sure was not created by AI,” according to Ahmed. “So, big tech companies then come to us and say, ‘We want 100,000 hours of content of people talking in the street.’ We go to our Anchor clients and say, ‘Do you have this?’” Anchor clients’ content is verifiable, he explains, because “we’ve already proven they don’t have AI in their archive.”\n\n*Anchor identifying media provenance (copyright OpenOrigins; used by permission)*\n\nAhmed references [C2PA, the open technical standard developed by the Coalition for Content Provenance and Authenticity](https://www.streamingmediaglobal.com/Articles/News/Featured-News/GL-CEO-Alexander-Leschinsky-Talks-C2PA-and-Todays-Digital-Content-Landscape-167514.aspx) for creating a verifiable record of content origin. While recognizing its value for content authentication in a world now teeming with deceptive AI-created content, Ahmed also acknowledges its limitations. “C2PA specifies how you show provenance, but it doesn’t tell you how to *establish* provenance,” he says. “Anchor establishes provenance. It proves that this content originated from this archive, and if you then want to take that information and export it as a C2PA file, we’ll let you do that.” \n\nThis information is useful to heads of archives, CTOs, chief information security officers, or anyone in a company who is in charge of its AI strategy and editorial. OpenOrigins does not currently offer public pricing. “The cost depends on whether we are doing lookups only on the visual scores or visual plus audio or if we are also creating a transcript and then doing a lookup on the transcript as well,” Ahmed says. “The rate at which you are expecting those lookups to happen, or the higher the rate of reads, the higher the cost.” As of late July 2026, the Anchor product has been in the market for more than a year.\n\n## **Videon**\n\n[Videon](https://videonlabs.com) is best known for its [LiveEdge encoding platform](https://www.videonlabs.com/liveedge-node) and for providing extensive “control from anywhere” for live, low-latency encoding. For this article, I’ll focus on [LiveEdge Cloud](https://www.videonlabs.com/liveedge-cloud-control), which delivers AI-driven video processing on its hardware, eliminating the overhead that comes from working with other LLM processing.\n\nLiveEdge Cloud leverages a neural processing unit (NPU) that is capable of providing more than 200 trillion operations per second. Users are able to run LLMs locally instead of utilizing a GPU-based solution in the cloud, which offers cost advantages. LiveEdge Cloud is also capable of running a number of AI models in parallel. This parallel operation “provides the mechanism for users to load different models” or applications, says Videon CTO and co-founder Paul Brown. “The platform allows the user to load, configure, and monitor the application(s).” Running multiple models in parallel “can enable a ball-tracking application to be run on the same audio/video feed as a speech-to-text application.”\n\n*Videon LiveEdge Platform*\n\nThe platform allows customers to supply more advanced-level delivery for both large and smaller events, as well as to provide more fan engagement such as social clips at the edge. Other types of things that can be detected within content are jersey numbers, identification of a person, or a bounding box. “With our system, the video and audio proceed directly from the input,” Brown notes. “Cloud solutions require transport and decode in the cloud. Depending on application type, [like] uploading only interesting content versus all content,” customers can realize “bandwidth savings of 90%.”\n\nThis cloud edge approach is also more energy efficient, says Brown. Nvidia and other cloud-based AI systems can use hundreds of watts of power, while Videon’s processing uses around 30 watts for comparable applications. In addition, AI local processing can generate data that may be used for other elements within the Videon workflow, such as creating SCTE-35 ad markers to identify where ad breaks should be and time-stamped metadata to align with the video stream.\n\n*Videon LiveEdge Cloud*\n\nLiveEdge Cloud has been in the market for more than a year, although the NPU is a recent addition. Its users include broadcasters, IP and streaming engineers, technical directors, and broadcast operations managers. No public pricing is currently available.\n\n## **Quickplay**\n\n[Quickplay](https://quickplay.com) is a Toronto-based company working to help its clients distribute content across broadcast, cable, OTT, FAST, social, and the creator economy. Its AI Studio is a web-based console that’s using AI to auto-create versions of content that may be trending with viewers. Quickplay VP of solutions engineering [Balbinder (Bal) O’Neil](https://www.linkedin.com/in/balbinder-o-neil-7b5329/) characterizes AI Studio’s mission as follows: “We need to be able to take that one piece of content and distribute it to multiple places, drive multiple formats, and then, importantly, plug and apply AI tools against the discrete problems sitting in the value chain.”\n\nAfter that content is imported into [AI Studio](https://quickplay.com/ai-studio-solution), O’Neil says, “I want to run an AI workflow to perform a scene-by-scene analysis on my video content. And this content could be live, or it could be VOD.” Scene and catalog information are surfaced via detailed metadata.\n\n*Promised time reductions and other efficiency improvements with Quickplay AI Studio*\n\nO’Neil provided a demo for me to show how metadata is used by the system to identify user queries. This is useful, she says, if “I want to create a very captivating moment about some content and instruct my AI editor to go look for content or for scenes where there are tense conversations happening in my content.”\n\nAnother example O’Neil offered was showing how more detail can be created. “Let’s say I had just ‘drama’ as my base genre, but it’s told me this [piece of content] is a thriller, a crime, a mystery. If this was sports, it would give me the teams and the actual players that are appearing in the content. I’ve now got hashtags, additional keywords, characters, actors. The LLM has actually given me different descriptions for different types of viewers: Gen Z’s, Gen X’s, or Millennials. I can now use all of this additional metadata for better search, for better discovery, and for better contextual advertising.” \n\nUsers can set business rules, such as “create clips between 30 and 60 seconds,” “exclude smoking or swearing,” or “don’t include the last 30 seconds because that’s a spoiler,” O’Neil explains. AI Studio further offers the ability to import overlays or graphics to use on top of this content. “What we can do in this screen can actually be imported into Adobe Premiere for additional editing capabilities,” O’Neil says. Or a user can do the edit directly within AI Studio.\n\nUsers can also change content formatting for social. “The LLM can determine what is important and auto-verticalize it,” O’Neil notes. The application’s Smart Crop feature “verticalizes content, making sure that the frame is always on the subject.” She showed me a demo where the vertical content within the frame switched from person to person, always keeping the active person in the frame.\n\nCropped or otherwise edited content can be sent to a CMS or social. It can be pushed out automatically or require user validation before distribution. According to O’Neil, current customers do both.\n\nQuickplay integrates with ad-decisioning servers like Google Ad Manager, FreeWheel, or SpringServe, with MediaTailor on the back end for the stitching. “We pass parameters,” O’Neil says, that “can be as basic as location and device type, based on what the user is actually watching contextually.”\n\nI asked O’Neil if there is a variable cost depending on how AI Studio uses models. “There is a cost on the type of model, and then there’s a cost on the processing time, number of hours processed, and the number of passes one might need to do,” she explains. “Sometimes, some models require you to do a couple of passes before you get enough metadata or enough relevance for the type of content that you’re pushing through.”\n\nA typical user of AI Studio would be a content marketer or product team. There is no public pricing available at present. AI Studio has been on the market since September 2025.\n\n            Related Articles\n            \n        \n\n                The three companies covered in this issue's column are all involved in bringing more monetization or engagement to today's streaming content. The key differentiator is using AI to delve deep into content meaning and adjacent data to create new opportunities for content formats that previously did not exist. While we sometimes cover tools that are suitable for smaller publishers, these AI tools require a certain scale of content to … well, work at scale. Two of these companies have been around for 10-plus years. The other one is much newer but already has an impressive connection with a media company.\n            \n\n                    23 Jun 2026\n\n \n                This is a tale of a few different AI use cases that involve research, localization, advertising, and UX. The first is a public broadcaster in Austria. The second is a TV OS. The third is a well-known vendor. The last is a major media company. What all of these have in common is that their AI applications have moved from the proof-of-concept stage to full commercial implementation.\n            \n\n                    26 Mar 2026\n\n \n                How has AI entered the media workflow? For this new column, we'll look at different applications used in the media industry. For this issue, we'll start with asset management, asset storefronts, and localization. While some of this functionality—speech-to-text transcription, translation, voice synthesis, natural language processing, logo detection, facial recognition, and object detection—has been around for a while, the biggest improvement is that much of it is now available on workflows with live content.\n            \n\n                    15 Dec 2025", "url": "https://wpnews.pro/news/ai-s-streaming-stack-image-interpretation", "canonical_source": "https://www.streamingmedia.com/Articles/ReadArticle.aspx?ArticleID=176364", "published_at": "2026-09-07 14:00:30+00:00", "updated_at": "2026-09-07 14:25:55.903381+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-tools"], "entities": ["OpenOrigins", "Manny Ahmed", "Anchor", "ITN"], "alternates": {"html": "https://wpnews.pro/news/ai-s-streaming-stack-image-interpretation", "markdown": "https://wpnews.pro/news/ai-s-streaming-stack-image-interpretation.md", "text": "https://wpnews.pro/news/ai-s-streaming-stack-image-interpretation.txt", "jsonld": "https://wpnews.pro/news/ai-s-streaming-stack-image-interpretation.jsonld"}}