Resonance™
Resonance is Adbustr's proprietary decision layer — a Golang-native, ClickHouse-backed inference system that turns every impression request into a conversion probability score in under 12 milliseconds.
Signal · Inference · Decision · Surface
The thesis
Most exchanges optimize for fill rate. The advanced ones optimize for eCPM. Resonance optimizes for the only number that matters to a performance advertiser: the probability of installing the app, completing the level, and returning on day seven.
We do this by training our models not on synthetic signals, but on the specific behavioral patterns of CTV-to-mobile attribution chains — a dataset class that simply doesn't exist outside our environment.
How it works
Adbustr SDK and server-side endpoints collect 200+ signals per impression: device class, app context, content metadata (per OpenRTB 2.6 Content object), behavioral entropy markers, supply path provenance.
Pre-bid scoring through ClickHouse materialized views runs in under 3ms. Impressions with IVT probability above 15% are dropped before reaching the auction layer — protecting our DSP partners from invalid inventory before they ever bid on it.
Resonance scores each surviving impression for predicted conversion value. The model considers historical performance of the placement, advertiser, vertical, time-of-day cohort, and contextual fit — outputting a continuous score that informs floor pricing.
Only the top-scoring candidate impressions are surfaced to DSPs at any given QPS budget. This reduces bid request waste by an estimated 60% versus traditional first-price auctions.
Performance lift
ROAS uplift
vs static auction
IVT exposure
for downstream DSPs
p99 latency
end-to-end
Numbers above are early-cohort averages from pilot integrations. Quarterly performance disclosures will be published in our Transparency Center.
Defensibility
Resonance isn't a wrapper around someone else's ML pipeline. The training data is proprietary to Adbustr's exchange position. The inference layer is co-located with auction infrastructure for sub-millisecond decisioning. The model architecture is engineered by a team that's spent five years optimizing for a single objective: turning CTV impressions into mobile installs that retain.