OpenAI Privacy Hackathon.
MVD helps developers reduce the data shared with AI by testing which information can be removed without changing the task’s outcome.
1st place
Less context. A clearer purpose.
Built as a team project at the OpenAI Privacy Hackathon at Station F in Paris, MVD is a skill for reducing the data sent to an AI task before processing begins. It compares complete records with smaller, purpose-specific inputs to find what can be withheld or replaced without changing the expected outcome.
From a task to an evidence trail.
Declare
Define the task, candidate fields, and expected decisions
Compare
Remove or transform inputs and rerun the same task
Constrain
Build a reviewed contract for the outgoing context
Explain
Record field decisions, observed changes, and limitations
From evidence to a usable report.
The prototype brings together Python evaluation scripts, an MCP server, and a React planning interface. My contributions focused on interactive privacy-minimization reports and HTML output from the certificate, making the field decisions and their supporting evidence easier to inspect.
A prototype with explicit limits.
The evaluation uses synthetic support tickets and a small BANKING77-derived test set. Findings apply to the declared task, data, prompt, and model; they are not a general guarantee. The planning interface produces a draft until an evaluation is run, and the resulting evidence supports an engineering review rather than certifying compliance.
OpenAI Privacy Hackathon participants at Station F, Paris.
