Bayesian predictive analytics for digital ad agencies

Forecast your campaigns before you launch them.

NoptiK gives ad agencies pre-launch revenue, conversion, and ROAS forecasts with calibrated uncertainty — so your forecasts are right about as often as they say they will be.

Posterior tightens as evidence arrives 2.5% median 97.5% prior posterior

Marketing forecasts are usually point estimates with no uncertainty.

Last-click attribution is broken. Marketing mix modeling is slow and retrospective. Both produce confident-sounding numbers that hide how much they're guessing.

Agencies running portfolios of clients sit on rich cross-client signal that today goes unused. NoptiK is built on a simple premise: forecasts should be honest about what they don't know, and the math should be auditable end-to-end.

Forecast at every stage of a campaign.

vertical channels budget audience Before the creative exists

Blueprint

Forecast from intent alone. Vertical, channel mix, budget, audience archetype. Used for new-business pitches and pre-creative planning. Outputs a three-tier band of likely outcomes.

scarcity trust urgency authority Before the campaign goes live

Sandbox

Forecast from a built campaign. The model reads the actual creative, breaks it into structured features (color, emotional valence, CTA framing, narrative arc), and produces a tighter prediction interval than blueprint.

regime shift As the campaign runs

Live

Sequential forecast that updates with each day's data. Intervals tighten as evidence accumulates. Regime shifts — platform algorithm changes, macro events, audience saturation — are detected automatically.

Watch a forecast tighten as a campaign runs.

Synthetic example: a fictional skincare brand running a 30-day campaign. Click through Blueprint, Sandbox, Live, and the calibration plot.

Demo data — synthetic example
Blueprint forecast · Heron & Pine · 30-day campaign $4,800 daily spend · Meta + Google · skincare vertical $28k $21k $14k $7k $0 Day 1 Day 15 Day 30 P50 (median) P10–P90 band
~$420k revenue
P50, 30-day total · Blueprint mode

Wide intervals because the model has only intent (vertical, channel mix, budget). The hive-mind contributes a prior from similar campaigns across the agency portfolio. Used for new-business pitches.

Sandbox forecast · Creative ingested · Features extracted Cialdini cues scarcity · 0.72 Plutchik tone trust · 0.61 $22k $16k $10k $0 Day 1 Day 15 Day 30
~$465k revenue
P50, 30-day · 32% tighter than Blueprint · Sandbox mode

The model has read the creative and extracted structured features. The interval narrows because the model now knows what kind of campaign it's predicting. The numerical forecast itself is still the Bayesian posterior — no LLM in the prediction loop.

Live forecast · Day 14 of 30 · Regime shift detected Day 9 $22k $16k $10k $0 Day 1 Day 14 (today) Day 30 Day 9 · regime shift realized · pre-shift realized · post-shift

14 days of realized data. BOCPD detected a regime shift on Day 9 (a hypothetical platform algorithm change). Rather than continuing to forecast off the pre-shift prior, the model widened the interval honestly. The forward fan from Day 14 reflects the post-shift world.

Calibration plot · synthetic illustration perfect calibration Predicted coverage Actual coverage 10% 30% 50% 70% 90% 10% 30% 50% 70% 90% When NoptiK says 80%, actual outcomes fall inside ~80% of the time.

The calibration plot is the headline accountability metric. Each point compares predicted coverage to actual coverage on a held-out window. Points on or near the diagonal mean the model's stated confidence matches reality. We commit to publishing this plot per-vertical, on rolling windows, in-product.

NoptiK is pre-launch and in active development. These visualizations illustrate the underlying methodology and forecast structure — the production interface, layouts, and output formats may differ from what's shown above.

Methodology, accountability, outcome. All three. All falsifiable.

01

Calibrated uncertainty

When NoptiK says 80%, we mean it. We measure whether our intervals actually contain the truth at the rate they claim, per vertical, on rolling windows. Coverage statistics will be published in-product when we go live.

02

No black box in the prediction loop

The forecast itself is pure math: Bayesian posterior plus conformal calibration. Language models extract structured features from creatives and write explanatory narratives — they never produce the numerical forecast. Reproducible from the audit trail.

03

Privacy by architecture

Each agency's data lives in its own isolated database schema with row-level security at the database layer. Cross-agency learning happens only through differentially-private aggregation — never on raw data.

Built by two co-founders. Equal partners.

K
Kevin Dehbashian
Co-CEO · President · Technical co-founder

Modeling, engineering, and inference. With Jonathan, previously built a macro predictive model that live-traded gold futures.

J
Jonathan Norarevian
Co-CEO · CFO · Operations co-founder

Operations, finance, legal, and go-to-market. Brought NoptiK's founding thesis; equal partner with Kevin on every product decision.

Read full bios →

A few things people ask first.

What does NoptiK actually do?

NoptiK is a Bayesian predictive analytics platform for digital ad agencies. It forecasts campaign outcomes — revenue, conversions, ROAS — with calibrated uncertainty, before campaigns launch and as they run.

How is this different from marketing mix modeling?

MMM is retrospective and slow. NoptiK is forward-looking and updates daily. Both use Bayesian methods; we differ in the prediction loop, the conformal calibration layer, and the cross-client hive-mind that learns across an agency's portfolio.

What about last-click attribution?

Last-click attribution is biased and broken in a post-iOS-14 world. NoptiK doesn't replace attribution — it forecasts outcomes directly, with calibrated intervals you can plan against.

Are you live yet?

Not yet. We're in active development with our first design partners. If you run a digital ad agency and want early access, apply to the design partner program.

What is the math behind NoptiK's forecasts?

Bayesian hierarchical models for partial pooling across clients, conformal prediction for distribution-free coverage guarantees, BOCPD for regime-shift detection, Postgres row-level security and differentially-private aggregation for tenant isolation.

Read the full FAQ →
We give you calibrated forecasts. When we say 90% chance, we mean it — and we'll prove it on your data before you sign.
Apply to design partner program → team@noptik.ai