Overview

Engagement Overview

Hark Audio partnered with TO THE NEW to reimagine its editorial workflow with an AI-driven engine that accelerates the discovery and curation of high-value podcast clips. Traditionally, editors spent hours manually scanning long-form audio to identify relevant moments. The new solution, built on a fusion of proprietary machine learning models and Generative AI, automates this process end-to-end. 

The platform intelligently identifies clip-worthy segments, refines boundaries with precision, and generates standardized metadata. By augmenting human editorial judgment with AI-powered insights, the solution enables editors to create richer, more engaging playlists in a fraction of the time while preserving their creative control.

Our Client

Media & Entertainment USA
Hark Audio is a U.S.-based media company specializing in curated podcast discovery. The company curates standout moments from thousands of episodes to deliver concise, theme-based playlists across news, culture, learning, and entertainment. Their mission is to simplify discovery and elevate the listening experience through expertly crafted editorial selections.
Rise

Business Objective

Hark Audio wanted to scale its editorial operations without compromising quality. The organization sought an AI-enhanced workflow that would:

01

Provide rapid, AI-generated suggestions of relevant podcast moments to reduce manual listening time

02

Enable editors to minimize effort with automated start–end boundary refinement for cleaner, more accurate clips

03

Maintain high-quality editorial standards with human-in-the-loop review and customizable refinement

04

Improve metadata quality with AI-generated titles, summaries, and tags to enhance searchability and organization

05

Increase editorial throughput by accelerating the discovery-to-publishing pipeline

06

Reduce operational cost by selecting and fine-tuning the most efficient model through benchmarking

Business Solutions

TO THE NEW designed and deployed a GenAI-powered editorial assistant embedded directly within the Hark Audio app, backed by AWS and robust intent modeling.

  • Trained a proprietary ML model to detect high-level clip-worthy segments across long-form podcast episodes
  • Fine-tuned the Nova Lite model on Hark’s internal dataset to produce precise start and end timestamps for each suggested clip
  • Integrated a dedicated LLM to generate clip titles, summaries, and tags, standardizing metadata across the library
  • Added a human-in-the-loop curation workflow so editors can review, select, and refine AI-generated suggestions
  • Leveraged Titan Embeddings and AWS OpenSearch to improve retrieval quality and ensure relevant clips surface quickly
  • Conducted model benchmarking across and outside AWS Bedrock, resulting in optimized cost, latency, and output quality

Business Outcomes

The collaboration enabled Hark Audio to fundamentally transform clip creation, elevating both productivity and content quality through an AI-augmented editorial process.

Faster clip discovery and generation through automated surfacing of relevant podcast moments

Higher editorial throughput enabled by AI-assisted boundary refinement and metadata generation

Improved accuracy of start/end timestamps via a custom-trained Nova Lite model

Reduced manual listening time, freeing editors to focus on creativity and curation

Consistent metadata quality through AI-generated summaries, titles, and tags

Optimized LLM cost after benchmarking multiple models inside and outside Bedrock

Tech Stack

  • AI Tech Stack
  • Vectorization
  • Retrieval & Search
AI Tech Stack
Vectorization
Retrieval & Search