SK Planet Shares Development Story of B2B Video Generation Service 'AI Moment' Combining GPT-4o and TTS

SK Planet revealed the development process of its self-developed AI-based video generation B2B service AI Moment. This service lets users simply enter a company name, after which GPT-4o classifies the category and generates marketing copy, and Google Cloud TTS synthesizes voice to produce the final video. It is used for corporate promotion in telecom value-added services such as V-Bizring.

Technical Challenges and Solution Strategies

The development team applied various techniques to address the non-determinism, exception handling, and speed issues of generative AI. To solve JSON format inconsistency problems, they enforced output format using Pydantic models and PromptTemplate. For exception situations such as TTS call failures, they secured service stability with a Graceful Degradation strategy that provides default copy.

Performance Optimization and Quality Management

To reduce delays caused by sequential processing, they applied Prefix Caching, cutting latency by up to 80% and cost by up to 50%. To guarantee the quality of AI-generated output, they introduced rule-based filtering to screen out inappropriate expressions and logic to adjust copy length. They also manage service quality through both qualitative evaluation (naturalness, brand suitability) and quantitative evaluation (failure rate, Latency, reuse rate).

SK Planet emphasizes that even when AI provides a probability-based 'valid answer,' it is the developer's role to judge it as a 'good answer' and connect it to user experience. They stressed that collaboration across the entire team, from prompt design to result structuring, is the core competitiveness of generative AI services.