← ResourcesTechnology5 min read

Two-tower bidding, explained

Lumero prices every impression with neural networks. Here is what that actually means, without the hand-waving.

The retrieval problem

A DSP sees enormous volumes of bid requests. For any one campaign, almost all of them are irrelevant. The first job of a bidding system isn't pricing — it's finding, in milliseconds, the small slice of supply worth pricing carefully.

Two towers, one similarity space

Our models are neural networks arranged two-tower style: one tower learns a representation of the campaign and its goal, the other learns a representation of the impression — app, placement, context, moment. Both map into the same vector space, where relevance becomes a similarity lookup. With ANN (approximate nearest neighbor) indexing, that lookup stays fast at auction scale.

Goal-aligned valuation

Retrieval narrows the field; valuation prices it. Every bid is priced to your true objective — install quality, ROAS, downstream events — not surface-level eCPM. If the goal changes, the pricing changes with it.

Continuous learning, honest statistics

Models update on live signal — wins, losses, conversions — so bidding sharpens as the campaign runs. And the system is statistically honest: decisions wait for enough signal to be real before the model acts, instead of overfitting to last hour's spike.

Inspectable, not opaque

None of this asks for faith. Where your budget went, what it bought, and how it performed are all visible in the same console our team uses. The transparency is the proof of the performance.