← All speakers

Supriya Subudhi

Conference affiliation: Staff Product Manager · Bumble · 2026

Supriya Subudhi is a product manager whose work at Bumble exposed a gap between winning an AI experiment and having a feature worth scaling. Working with the team behind Profile Guidance and AI Photo Feedback, she confronted three different results: people said they wanted advice, relatively few used it, and advice that prompted profile edits could still be wrong.

On this page

From engineering to consumer products

Supriya worked as a software and data engineer at Microsoft before studying for an MBA at the Indian School of Business from 2018 to 2019. Her product career subsequently included advertising technology and fintech, including PayPal, before she joined Bumble in London. Her professional history spans both engineering and responsibility for consumer products; she was credited as a staff product manager at Bumble in the 2026 AI Engineer NYC roster.

At Bumble, her team addressed a specific problem: members could assemble a dating profile without knowing whether it represented them well. Profile Guidance offers feedback on written bios and prompts, while AI Photo Feedback suggests improvements to photo selections. Bumble announced the features in February 2026, with written guidance rolling out globally and photo feedback available in the United States.

What a winning experiment leaves unanswered

In her account of building AI photo feedback, Supriya describes how encouraging research and positive experimental results left important product decisions unresolved. The experience changed what she expected an A/B test to establish: measured user benefit still needed to survive scrutiny of the advice, its delivery and its cost.

  • Put advice where people are already acting. Roughly 78% of research participants welcomed profile feedback, but only 10–15% of people offered the experimental feature used it. Adoption improved when the team integrated advice into existing activities, such as uploading photos. Interest in improving a profile did not mean members would seek out a separate feedback feature.
  • Read the advice behind the metric. More profile edits did not establish that the recommendations were useful. Some members who were already smiling were told to smile more. Supriya argues for inspecting individual outputs and evaluating how advice differs across user groups. In her AI Engineer presentation, she also identifies concrete limits: feedback should not ask people to change their height or skin color. These checks address failures an aggregate experiment result can conceal.
  • Test the model that can actually ship. The stronger model produced results the team could not afford to scale. Cheaper models and conventional machine learning reduced costs but weakened the impact. That tradeoff made model selection part of the product experiment: a result measured with one system could not establish the value of a cheaper replacement without another test.

Supriya continues to advocate randomized experiments. Her changed position concerns the launch decision: a positive result must hold for the advice people receive, the model the company can afford and the markets where the feature can be offered.

1 conference talk

Loading more talks and highlights…

References