Frédéric Petitpont Newsbridge CTO & co-founder
What was the biggest technology challenge you’ve had to face in your work? The biggest – and most rewarding – has been to lead a team in the creation of customised artificial intelligence pipelines. These make the indexing of audiovisual content incredibly accurate and affordable for companies wanting to preserve content forever. Newsbridge’s mission is to help organisations immediately access everything they’ve ever produced. We promise that our users will be able to find the photo or video clip they need in less than two seconds, whatever the size of their dataset. We could be talking about a century of football club archives, 200,000 hours of video or 40 live, simultaneous video streams. To achieve this, we use automated indexing AI. Existing AI technology for indexing video came with many limitations, like biases (for example, misdetecting faces), transcription errors and, above all else, incredibly high running costs. Our own AI pipelines had to challenge the market standard. It’s becoming less and less acceptable in many organisations to spend eight hours looking for specific content on tape or through external hard drives, when it suddenly becomes relevant again. The traditional way of managing, storing and indexing content simply doesn’t fit the time constraints of today’s world. And I’m not even talking about teams working remotely that can’t access their LTO library! What were the stakes for not resolving it? Every two years, the volume of content produced worldwide doubles. More media made equals more stored. It’s really hard for a company to make the conscious decision to hit the delete button. It has become easier and cheaper to store everything, rather than trying to sort and predict what you will use again. But storing content without proper indexing means it’s basically lost in the abyss. Over time, we’re even talking about a loss of collective memory.
STORING CONTENT WITHOUT PROPER INDEXING MEANS IT’S IN THE ABYSS
What happened? How did it work out? We had to make indexing AI reliable and affordable, to handle the endless accumulation of content. To fight AI biases and boost quality, we made it multimodal. This is based on data fusion. We are merging together detection results from face, text, objects, pattern and transcription. All of this contextual information produces incredible metadata and search results. Addressing the affordability issue, we designed our AI to be more energy- efficient. We began to use more CPU- based filters at the start of each pipeline (such as scene sampling, subsampling, tracking and audio analysis), to reduce inferences and therefore use less GPU. The result was greatly reduced energy consumption, cost and processing time, but also a lower carbon footprint. We now have thousands of terabytes of photo and video running through these customised pipelines every month. What did you learn – and what would you do differently if you had your time again? So, if I could talk to the CTO I was four years ago, I would probably give him the set of KPIs we work with at Newsbridge today. To scale a tech team with a culture based on trust and horizontal collaboration, KPIs should never be underestimated. Everyone needs a North Star. Setting clear and challenging energy, quality and running cost targets maximises everyone’s efforts, improves team velocity and personal fulfilment.
@feedzinesocial
Powered by FlippingBook