How Acast hacked its data

How Acast hacked its data

Or "building fun stuff without pressure"

Written by Ioana Havsfrid, Richard Jenkins2024.12.16

At Acast we want to be pleasantly surprised by what we and our technology can do. Some breakthrough ideas are harder to fit into a product roadmap so we tap into other ways to play with them. Our preferred way to do this are our twice-yearly Hackathons! These are weeks dedicated to trying out new things, building cool proofs-of-concept, and learning from colleagues you may not usually interact with on a daily basis.

How hackathons work at Acast

High-risk / high-reward ideas using the shiniest new tech sound great, but implementing them is rarely easy. We believe that constraints drive creativity, and therefore each hackathon has a theme. 

This time it was all about data! Not an original theme, we know, but it made sense in context! Over the last 10 years Acast has become a podcasting oracle, sitting on vast amounts of data across everything from production, distribution, and consumption to trends and flops. This huge pool of data gives us opportunities but can also overwhelm us. For this reason we decided to dedicate our hackathon to finding new ways to use, display, and process our data with no external pressure to hit a deadline or release something tangible.

We were pleased this time to receive some special assistance from our friends at Amazon Web Services. Niklas Palm, Ebtisam Mohammedsalih, Muhammad Sajid, and Heidi Sollheim visited our Stockholm office to help one lucky hack team plan the architecture for a machine learning project focused around podcast discovery & media planning, code-named “RossBot”.

Niklas explains a retrieval-augmented generation reference architecture to the group

At the end of each Hackathon, the department comes together to watch the demo session! Each hack team demonstrates what they’ve been working on - that could be in the form of a proof of concept, or simply a presentation explaining “we tried this, and learned a way not to build it”. Learning is the key thing, rather than tangible outputs.

The demo session for autumn 2024, including the all important fika provided in the office!

This year’s highlights

‘RossBot’ - a robotic planning pal!

In order to keep the world’s most valuable podcast marketplace running, matchmaking advertisers with creators, our sales teams naturally spend lots of time finding the best shows for their clients. Usually this involves finding those that align with the audience an advertiser wants to reach - such as “young women living in New England who play football”.

This process can be tricky, as although we have a wealth of data available on shows it can be hard to sift and sort that data to pick out the shows you need. Enter RossBot! Named after our CEO Ross Adams and powered by a Retrieval-Augmented Generation model, it serves up the shows best aligned with the user’s query in less than 30 seconds. Most importantly it also explains why the selected shows fit the brief in natural language that a seller or their client would understand.

We believe these sort of proactive recommendations could be transformative for planning, and aim to iterate on this project in the near future. Watch this space…

Social media data downloads for our sales teams

Podcasters were the original influencers, and many podcasters now also create content on Instagram, TikTok, and other platforms. This offers advertisers a powerful opportunity to reach those creators’ audiences in an omni-channel fashion, across multiple platforms.

Doing this of course requires our wonderful sales teams to have data on those audiences - size, demographics, engagement.

The product team that looks after our ad campaign planning & booking products decided to use the hackathon to supercharge this initiative! Their aim was to display as much data as possible about our creators’ social media channels in our internal planning suite.

The team combined data from Podchaser with some additional points from external tools they explored. Most impressively they built data pipelines, a backend, and a fully-designed frontend UI to make this information accessible to users during the week. A shining case of ‘move fast and build things’!