Custom models & RL environments

An ad optimization model trained on your business.

Mindra trains its own models for ad decisions. For enterprises, we post-train them on your data, build RL environments from your accounts, or train a model that is entirely yours.

Updated

In short

Mindra's budget and bid decisions come from ad optimization models it trains itself. As a separate service for enterprises, Mindra post-trains those models on a company's own data, builds reinforcement learning environments from its ad accounts so policies can be trained and tested without risking live budget, or trains fully custom models for decisions specific to that business. Models go live through Mindra's agents, rules and guardrails, after a backtest and a shadow period.

Best for
Enterprises spending $1M+ a month, or with in-house data science
Delivered as
Post-trained model, RL environment or fully custom model
Your data
Trains your model only, never pooled
Pricing
Scoped per engagement, on a call

The service

Three ways to work with us

  • Post-trained Mindra model

    Our base ad optimization model, post-trained on your account history, margins, seasonality and market mix, so it learns how your business converts rather than how the average advertiser does.

  • RL environment

    A simulation environment built from your ad accounts and revenue data, with your business goal as the reward. Train and evaluate allocation and bidding policies, ours or your own team's, without risking live budget.

  • Fully custom model

    A model trained for a decision specific to you, such as allocation across retail and marketplaces or bidding on a margin target, deployed through Mindra's agents, rules and guardrails.

Why it matters

Why a model of your own

  • General models learn the average

    A model trained on everyone's data learns the average advertiser. The places your markets, margins and channels differ are exactly where the money is.

  • Live budgets are a costly classroom

    Testing a new policy on real spend is slow and expensive. An RL environment lets policies learn and be compared on your own dynamics first.

  • Your goal, not a proxy

    Contribution margin, payback or qualified pipeline become the objective the model optimises, instead of clicks or platform-reported conversions.

Read how Mindra's ad optimization models work, or see the full enterprise offer.

How an engagement works

Proof before it touches a live budget

  1. 1

    Data audit

    We map your ad accounts, store, payments, MMP or CRM, margins and anything in the warehouse.

  2. 2

    Build the model or environment

    We post-train the model, build the RL environment from your history, or train the custom model.

  3. 3

    Backtest on your history

    We replay past periods and show, day by day, what the model would have done and what it would have returned.

  4. 4

    Shadow mode

    The model recommends while your team keeps running the budget, until you trust its calls.

  5. 5

    Go live under your rules

    It starts acting inside the limits and approval thresholds you set.

  6. 6

    Ongoing tuning

    Our engineers retrain on new data, adjust for launches and seasonality, and review results with you.

Your data, your model

What the contract guarantees

  • Your data trains your model only. It is never pooled into a shared model or used for another customer
  • Tuned weights are yours for the life of the engagement, and the contract says so
  • SOC 2 Type II and GDPR, with Zero Data Retention and EU data residency options
  • Forward-deployed engineers and a named team accountable for results

Custom models FAQ

What is an RL environment for advertising?

A simulation of your advertising built from your own history: how spend, bids and creative choices in each channel and market turn into results. Reinforcement learning policies can be trained and evaluated inside it, with your business goal as the reward, before they touch a live budget.

Does Mindra train its own models?

Yes. Mindra's budget and bid decisions come from ad optimization models we train ourselves. This service extends them to a single company's data, or builds new ones for it.

Who owns the model?

Your data trains your model only and is never pooled or used for another customer. The tuned weights are yours for the life of the engagement, as stated in the contract.

Can our own ML team use the RL environment?

Yes. The environment can be used to train and evaluate your team's own policies as well as Mindra's.

Who is this for?

Typically brands, apps and platforms spending $1M+ a month across several channels and markets, or teams with their own data science function. Pricing is scoped per engagement; book a call to discuss your setup.

Still have a question? Write to info@mindra.co and we'll get back to you.

Start with a backtest on your own numbers.

Tell us about your markets, margins and channels. We will show you what a model trained on your data would have done, before anything touches a live budget.