64950 - Data Analyst

Data Analyst

Location: London, Hybrid onsite Brock Street - 3 days per week 
Start date: ASAP 
Duration: 12 Months
Minimum Pay Rate: £25.00 per hour (Inside IR35)


Despite rapid development in generative AI, quality measurement remains challenging due to the lack of reliable and robust technical metrics. This is especially the case when the slight differences in the details generated by the model could have a significant impact on user experience. In this team, we are developing a series of manual (human rating), semi-automatic, and fully automatic evaluations to measure the quality of our image and video generative models (text to image, image to video, text to video, etc.) in a variety of quality axis, including visual quality, prompt following, identity preservation, naturalness, and visual text generation and so on.

Job Responsibilities:
Working with the team to prototype human rating and semi-automatic model evaluation protocols.
For high stake and / or high complexity protocols, conduct small-to-medium scale in-house human rating on generated content.
Working with engineers to identify and summarize model gaps based on the above evaluation
Working with engineers to identify data needed to close those gaps and mine such data to prepare training data for the next iteration of model training
Working with PDO teams to scale up validated evaluation protocols, including coordination and auditing.

Skills:
Verbal and written communication skills, problem solving skills, and interpersonal skills.
Attention to details and an aptitude to experimental investigations
Basic ability to work independently and manage one’s time.
Basic knowledge of Python, Unix, and SQL.
Basic knowledge of computer vision and generative models

Education/Experience:
Associate's degree or equivalent training required in Computer Science, Electronic Engineering, Physics, Bioinformatics, or other STEM subjects.
Prior industrial experience in software development and testing and / or research experience in human computer interaction are preferred.