Skip to content

How it works. Rehearsed before the bell.

Hi Violet trains its own models with reinforcement learning on more than 10 TB of data. Social models are tuned per niche to script and edit for hook and watch time. AI agents work through your pages and software in our own RL environments. Ad variants face AI-simulated focus groups before any spend.

See every service
of data behind our models
10+ TB
gaming and Web3 niches trained
50+
platforms: TikTok, X, YouTube Shorts, Reddit
4
Seen from behind, a floor manager in a violet jacket holds up a stopwatch at the rail while traders in grey practice jackets rehearse hand signals in the empty pit below.

One loop under every move

  1. 01

    Generate

    Two dozen variants of one post, page, or ad.

  2. 02

    Simulate

    A simulated audience or an agent run scores every variant.

  3. 03

    Reward

    Reinforcement learning rewards the top scorers.

  4. 04

    Ship

    The strongest variant goes live.

  5. 05

    Retrain

    Live results become training data for the next round.

Every post, page, and ad starts as a field of variants. Simulation scores each one, reinforcement learning rewards the winners, the strongest variant ships, and its live results become training data for the next round.

Social models trained on the niche.

We fine-tune our own models with reinforcement learning across more than 50 niches inside gaming and Web3, in short video, static images, and text. Each model learns what holds attention in its niche, then scripts and edits for it, so a post reads like it came from inside the community.

What the models are rewarded for

  1. The hook

    The first one to three seconds of a clip, the first frame of an image, the first line of a post. The reward is attention held past the opening.

  2. Watch time

    Pacing, cut timing, captions, and where the payoff lands, scored on retention, completion, and rewatches.

  3. Niche voice

    The references, slang, and formats each niche trusts, learned from the niche itself, so nothing reads as generic.

  4. Platform fit

    One idea written natively for TikTok, X, YouTube Shorts, and Reddit: vertical edits, threads, thumbnails, and long-form posts.

Hook and hold

  1. 01Hook window
  2. 02The hold
  3. 03Payoff
  • Generic edit
  • Niche-trained edit

Illustrative retention curves for one 30-second gaming clip. The niche-trained edit keeps more viewers through the first three seconds, holds them through the middle, and lifts again at the payoff, where rewatches start.

Agents learn your product first.

We rebuild your docs, signup, and API inside our own reinforcement learning environments. AI agents attempt the tasks your users hand them, run after run. A finished task earns the reward, every stall marks a page to fix, and page and docs variants compete until agents finish reliably.

Inside the environment

  1. A copy of the route

    Your docs, signup, quickstart, and API, rebuilt in a sandbox, so agents never touch production data or real money.

  2. Repeated runs

    Coding agents, browser agents, and assistants run every task again and again, on current models.

  3. A reward for finishing

    A run scores only when the job is done: an account made, a key issued, a first API call returned.

  4. Variants until it holds

    Page, docs, and error-message variants compete in the environment. The version agents finish most often ships to your real site.

Runs that finish, round by round

  1. Round 14/20
  2. Round 27/20
  3. Round 310/20
  4. Round 413/20
  5. Round 516/20
  6. Round 618/20
  • Finished the task
  • Stalled or quit

Illustrative environment for one task, twenty agent runs a round. Each round ships the page and docs variants agents finished most often, and the next round runs against them.

The panel sees it first.

Before an ad spends a dollar, we generate its variants and put them in front of AI-simulated focus groups, run in aggregate at scale. The panels react to every variant, reinforcement learning tunes the generator toward what they reward, and only the strongest variants go live.

What the panel scores

  1. The hook

    Whether the opening stops the scroll, and for which audience.

  2. The CTA

    Whether the ask is clear, believable, and worth acting on.

  3. The gesture

    The small cues that point at the action: a glance toward the button, a hand toward the link, a pause before the offer.

  4. The implied meaning

    What the ad says without saying it, and whether that reading helps or hurts the brand.

From behind, a row of observers in a dim gallery watches a packed trading pit through the glass, one resting a pencil on a blank card.
The panel watches before the market does

Six variants, one panel

  1. Variant A
  2. Variant Bgoes live
  3. Variant C
  4. Variant D
  5. Variant Egoes live
  6. Variant F

Scrolls pastActs on the CTA

Illustrative panel for six variants of one ad. Each dot is one simulated panelist. The two variants the panel rewards go live; the rest are rewritten or cut.

Where the models sit in the cycle.

  1. Read

    The trend engine's 50K+ weekly signals and your own results become fresh training data for each niche.

  2. Choose

    Simulated panels and agent runs test the candidate bets before one is chosen and written into scope.

  3. Ship

    The trained models script, edit, and generate the posts, pages, and ads that go out.

  4. Measure

    Live retention, clicks, and finished agent runs score the models again, so each cycle starts sharper.

Where the models run

Questions buyers ask

Does Hi Violet train its own AI models?
Yes. We fine-tune our own models with reinforcement learning on more than 10 TB of data, for social scripting and editing, agent experience, and ad testing.
What is reinforcement learning in marketing?
Training a model by reward. The model makes a hook, an edit, an ad, or a page change, gets scored on the outcome that matters, such as watch time, a finished agent task, or a panel's response, and shifts toward what scored well.
Which niches are your social models trained on?
More than 50 niches inside gaming and Web3, in short video, static images, and text, for TikTok, X, YouTube Shorts, and Reddit.
What is an AI-simulated focus group?
A panel of simulated viewers that reacts to ad variants before launch. Run in aggregate at scale, the panels score the hook, the CTA, the gestures that point at it, and what the ad implies, so spend starts on the variants most likely to work.
What is an RL environment for AI agents?
A sandbox copy of your docs, signup, and API where agents attempt real tasks over and over. A finished task earns the reward, and every stall points at the page, doc, or error message to fix.
Do simulations replace live testing?
They come first. Simulation narrows the field before spend; live results then score the models again, so every round of simulation tracks real audiences more closely.
Who owns the content the models make?
You do. Posts, edits, ads, and page changes made for you belong to you, like every other deliverable in a Hi Violet scope.

Rehearse it before you spend it.

Tell us what you want to move. You get a written scope, with every deliverable and price, before any work starts.

See engagements and pricing