FDEBench Join us live

Q
8 / 159

Multimodal & voice

How do you design a voice AI that retains user preferences and conversations while ensuring data can be deleted in under 1 second upon user request?

question bank FDEBench bank Source

01
Ask

What a strong candidate asks first

Before designing anything. Skipping this is the most common way the round is lost.

  • Does "deleted" mean unrecoverable, or unreachable-and-eventually-purged?
  • Does the requirement cover backups, logs, and derived data?
  • Is one second the user-visible confirmation, or actual erasure?

02
Rubric

What earns the points

6 criteria, 100 points total. Scored independently, partial credit allowed.

Maps everywhere user data lives 25 pts

Maps every location user data reaches — primary store, caches, search/vector indexes, logs, analytics, backups, training sets. Three or more locations is full credit.

If you miss it You deleted from the database. User data is also in caches, indexes, logs, analytics, and backups — a deletion design is mostly an inventory problem.
Deletes the key, not every copy 25 pts

Recognizes that traversing every store in one second is impossible at scale and uses indirection instead — per-user encryption keys (crypto-shredding), a tombstone that gates all reads, or equivalent.

If you miss it You tried to physically delete from every store inside one second. That doesn't scale. Use indirection — make the data unreadable instantly by destroying a per-user key or gating every read on a deletion marker — then purge in the background.
Hidden now vs erased for good 15 pts

Separates "immediately unreachable" from "eventually physically erased", and states a target for the second.

If you miss it You conflated instant and permanent. Say plainly which one your design delivers in a second and how long the actual purge takes.
Handles backups and archives 15 pts

Addresses backups and immutable archives specifically, rather than ignoring them.

If you miss it You didn't address backups. They're the hardest part — you usually can't edit them, which is a strong argument for the encryption-key approach.
Covers embeddings and fine-tunes 10 pts

Raises data derived from the user — embeddings, model fine-tunes, aggregate features — and what deletion means for it.

If you miss it You deleted the raw data but not what was derived from it. Embeddings and fine-tuned weights carry the user's information forward.
Can prove the deletion happened 10 pts

Can demonstrate the deletion happened — audit record, verification job, or user-visible confirmation.

If you miss it You can't prove the deletion occurred. Under a privacy regulation that proof is the deliverable, not the deletion itself.

03
Flags

Red flags

Reported separately, not subtracted — this is the "what would have ended the interview" note.

  • Proposes a synchronous cascading delete across all stores.
  • Never mentions backups.

—