Model Architecture & Intelligence
The AI behind every bidding decision.
Real-time advertising decisioning requires specialized models working together in parallel. Rather than treating machine learning as an opaque magic box, we share our component architecture, training methodology, and current deployment status transparently.
Model Components
Dedicated models. Calibrated outputs. Verified status.
Under our Accuracy Rule, every machine learning capability is explicitly labeled with its verified production state: Deployed or In Partner Testing.
Click-Through Rate (CTR) Estimation Model
Accurately forecasting whether a given user environment and ad format will generate an engagement before committing budget.
Domain history, ad placement dimensions, operating system, dayparting hour, and contextual category taxonomy.
Calibrated probability score (pCTR between 0.0001 and 1.0) utilized in CPM valuation formulas.
Gradient-boosted decision trees (GBDT) coupled with logistic regression feature interactions.
Conversion & Action Likelihood Scoring
High CTR does not guarantee high post-click conversion rates. Bids must reflect true business value (CPA / ROAS), not empty clicks.
Historical campaign conversion signals, landing page category alignment, connection type, and device type.
Predicted conversion likelihood (pCVR) dynamically modifying the target bid floor.
Deep factorized neural embeddings evaluated via lightweight server-side inference.
Invalid Traffic (IVT) & Ad Fraud Detection
Ad fraud, non-human bot traffic, and spoofed domains drain advertiser budget and skew optimization metrics.
Known datacenter IP registries, anomalous user-agent patterns, historical bid-stream velocity, and domain authenticity checks.
Pre-bid fraud score (0 to 100); opportunities exceeding risk thresholds are dropped prior to bidding.
Heuristic rules paired with isolation forests and real-time IP reputation lists.
Continuous Self-Learning Optimizer
Market clearing prices fluctuate throughout the day. Static models quickly become obsolete as competitor bids shift.
Streamed auction win/loss telemetry, clearing prices, and impression viewability signals.
Adaptive bid multipliers and updated prior weights for upcoming auction requests.
Reinforcement feedback algorithms and multi-armed bandit exploration/exploitation loops.
Natural-Language Rule Parser & Compiler
Translating human business goals ('increase bids during evening rush') into mathematically sound, low-latency execution constraints.
Advertiser free-text prompt input and current campaign parameters.
Structured JSON condition tree with human approval verification and safety check report.
Domain-fine-tuned semantic parsing model with deterministic logic validation layers.
Explainable AI Commitment
Autonomy with a ceiling, never a black box.
In programmatic advertising, runaway models without human boundaries can exhaust entire campaign budgets on misleading signals. Visprisca binds all autonomous learning within explicit user-defined guardrails.