Hirundo
Machine unlearning: removes the influence of specific bad or unlawful training data from a model without retraining it from scratch.
Highlights
- Machine unlearning platform that removes bad data influence from trained models without full retraining
- Founded in 2023 by Ben Luria (ex-IBM Research) and Roi Reichart (Technion professor)
- Headquartered in Tel Aviv, Israel
- Approximately $8M+ seed round; backed by Maverick Ventures Israel and Magma Venture Partners
- Used for GDPR right-to-be-forgotten, copyright removal, bias remediation, and PII purges
- Targets foundation models, fine-tuned LLMs, and traditional ML models
- One of the few commercial vendors in the emerging machine-unlearning category
External link — opens hirundo.io in a new tab. Hirundo is a third-party product; we are not affiliated with it.
About Hirundo
What it is
Hirundo removes the effect of particular training data from an already-trained model — bad labels, copyrighted material, personal data subject to a deletion request, sources that introduced bias — without the full retraining that would otherwise be the only remedy.
Why it's different
This addresses a problem that has no other practical answer. A foundation model cannot reasonably be retrained because somebody exercised a right to erasure, or because one dataset turned out to be unlicensed, and "we cannot remove it" is increasingly not an acceptable reply to a regulator. Unlearning is the only route that is not ruinously expensive. The caveat is that it is an active research area rather than a solved one: verifying that influence has genuinely gone, rather than merely become hard to elicit, is itself unsettled, and a compliance claim resting on it is resting on a young technique. Ask what evidence of removal looks like.
How people use it
It is used by teams who have discovered something in their training data after the fact — a licensing problem, a privacy request, a source that poisoned a behaviour — and need it gone without starting again. The question worth pressing in any evaluation is how removal is demonstrated, because the value of the whole exercise is the proof rather than the operation.
Written by the n3os team. We are not affiliated with Hirundo.
This listing was written from public information, without Hirundo’s involvement. If you own it and something here is wrong — or you would rather not be listed at all — email us and we will correct or remove it.
Get the ones worth knowing about
We write one of these for every tool worth the trouble. Get the new ones, plus what we have found genuinely useful lately.
Your address goes to Buttondown, who send the emails on our behalf. One click unsubscribes, and the list is never sold or shared.