How NoptiK forecasts campaigns.
Three modes — Blueprint, Sandbox, Live — covering every stage of a campaign from new-business pitch to mid-flight optimization. One Bayesian model underneath. Cross-client learning while keeping every tenant's data strictly its own.
Forecast before a single dollar is spent.
Blueprint mode answers a question every agency hears at new-business pitches: "If we ran this campaign, what would happen?" With no creative yet, NoptiK draws on the hive-mind alone.
The prior comes from what similar campaigns have done across the agency's portfolio: similar vertical, similar channel mix, similar budget envelope. The model returns a three-tier band of outcomes — Conservative, Expected, Aggressive — with calibrated probability around each tier.
Bands are wide because the data is sparse. They tighten in Sandbox mode once a creative artifact exists. Used for new-business pitches, pre-creative planning, quarterly portfolio forecasting.
Forecast before the campaign goes live.
Once the creative exists, the model reads it. Vision and language models extract structured features: color palette, emotional valence (Plutchik wheel), persuasion cues (Cialdini's reciprocity, social proof, scarcity, authority, liking, commitment), CTA framing, narrative arc.
These features feed the prior, sharpening it. The forecast tightens because the model now knows what kind of campaign it's predicting.
Critical distinction: the language model never produces the numerical forecast. It only extracts features. The forecast itself is the Bayesian posterior, conditional on those features. Every feature extraction is logged and reproducible.
Used for go/no-go decisions, budget allocation, creative A/B/C/D comparison before launch.
Forecast as the campaign runs.
Each day's platform data updates the posterior. Yesterday's forecast becomes today's prior, weighted by measurement quality — noisy days contribute less, clean days contribute more. The forecast tightens with evidence.
A Bayesian online change-point detector (BOCPD) flags regime shifts: platform algorithm changes, macro events, audience saturation, creative fatigue. When a regime shift is detected, NoptiK widens the interval honestly rather than continuing to forecast off stale priors.
This is the difference between a model that pretends nothing happened and a model that admits the world has changed.
Used for in-flight optimization, daily pacing decisions, automated alerts on anomaly detection.
A hive-mind that respects tenant boundaries.
Every agency's data lives in its own isolated Postgres schema with row-level security enforcement at the database layer. There is no cross-tenant query path.
The hive-mind doesn't pool raw observations — it pools aggregated posterior parameters that have passed through a differentially-private aggregation layer. Mathematically, no individual client's data can be reconstructed from what gets shared. Auditable end-to-end.