Adaptive

Adaptive Security trains companies to spot scams made with AI. Its sales team won customers over with a deepfake video of their own CEO. Customers then found they couldn’t make one themselves. I designed the tool that lets them.

Role
Product designer. Design, prototypes and testing the AI models.
Worked with
Engineering, the team that made deepfakes by hand, and marketing.
Built with
ElevenLabs for the voice. Veed, Runway and HeyGen tested for video.
A page in the Adaptive app for one persona, Jared Long. A box to type a video script sits above a library of the audio and video clips already made with his face and voice.
A customer types a script in the box at the top and gets a video of this person saying it. The clips already made are listed underneath.

The problem

Adaptive already let a security team clone their CEO’s voice and use it in fake phishing calls and training. In sales calls the team showed more than that, including deepfake video made by hand for each prospect.

This is a deepfake of me, made from one photo and 30 seconds of audio to demonstrate the technology.

Customers loved it in the demo. Once they were in the product they found a page of voice settings and no way to make a video. It was one of the biggest complaints from new customers, and it could derail a deal during the trial.

Deepfake video of your executive

In the product, made for you by Adaptive.

Deepfake phone call you can talk to

In the product.

Deepfake audio clips you make yourself

In the product.

Deepfake videos you make yourself

Shown in the demo. Not in the product.

Deepfake video you can talk to

Shown in the demo. Not in the product.

The old persona page for a persona called Michael Scott. Most of the page is sliders for the voice: stability, speed, similarity and style exaggeration.
Before the redesign, most of the page was voice sliders and there was nowhere to make a video.

What customers asked for

We talked to five customers who had used the persona features. They asked for three things.

Manage it themselves

Find, make and change their deepfake videos without sending Adaptive a request.

Themed narrators

Celebrity or themed presenters for security training, so employees pay attention.

Live video

A persona they could put on a video call and talk to.

The first one was the complaint, so the first release was about that.

Choosing a model

The people inside Adaptive who made deepfakes had tried many models, and often built a one-off process to get a single good result. A customer can’t do that. For the tool to work without help I had to test the models and pick the one that failed least.

Customers rejected anything that broke the illusion. A wrong background or an odd gesture and they said “I don’t buy it,” and the video was no use for training. They were just as hard on likeness. If the face wasn’t exactly the person, they said so.

Veed

Kept the likeness best. A little stiff.

Runway

The most expressive, with the most movement in the scene.

HeyGen

Lifelike talking heads.

I kept the tool narrow on purpose. You upload a photo and write a script. It speaks the script in the person’s voice and moves the face in the photo to match. It doesn’t invent a scene or take open-ended prompts, because customers cared more about a believable face than about choosing the scene.

What I changed

Hid the settings

The voice sliders moved into a one-time setup. Most customers never need them again.

Put the videos first

A persona’s page became a library of what has been made, with thumbnails, dates and the first line of each script.

No naming

Nobody was going to write useful names, and the list would have filled with “untitled.” The tool keeps things in order.

No steps

Type a script, make the video, change it, make it again. Everything starts on a sensible setting.

The AI Personas page in the Adaptive app, with a card for each person a company has cloned and a count of the videos and audio clips made with each.
A company sees everyone it has cloned on one page, with a count of the videos and audio clips made of each.

What happened

Customers who felt they had already paid for this got it, and the complaint stopped.

Not many people use it, and that was always going to be true. At each customer, between one and ten people work in the product at all. A few of them use this tool heavily.

The reps who book first meetings and the account executives who close deals both told me it helped them sell.

Inside Adaptive, the team stopped making deepfakes by hand. New accounts now start with the photo and voice already in place.

The photo booth

Customers kept asking to put the person somewhere other than the room in their photo. That was out of scope for the first release, so I built it on the side: describe a place, and the person appears there.

A generated picture of me standing in a misty jungle with two dinosaurs behind me. A generated picture of my face inside an astronaut’s helmet, with the Earth behind. A generated picture of me in a linen shirt on an overcast beach.
The same photo of me placed in a jungle, in space and on a beach, each from a written prompt.
One of the prompts

Photorealistic cinematic portrait of a person inside Jurassic Park. The subject is framed from mid-torso upward, centered in the frame. Eye-level camera angle, natural cinematic photography perspective. The subject stands calmly in the foreground, alert and composed, subtle tension in posture. Looking directly at the camera or slightly past it, focused expression. Background shows a lush prehistoric jungle environment — tropical plants, mist, rain-soaked atmosphere — with large dinosaur shapes or silhouettes visible in the distance, partially obscured by foliage. Soft, diffused jungle lighting with humidity and mist in the air, no harsh shadows. Shallow depth of field with the subject sharp and the jungle environment layered behind them.

A workflow in Fal made of four connected boxes: the inputs, a model that draws the person in a new scene, a model that moves the face to match the audio, and the finished video.
The workflow in Fal uses two models. One draws the person in the new scene and the other moves the face to match the voice.

Then the CEO asked for a deepfake people could try for themselves at events and trade shows. I built a proof of concept on top of the side project to show the team it could be done. There was no product manager on it and engineering was stretched, so I designed the whole thing and shipped it with marketing.

At an event, a visitor takes a selfie, reads a line aloud and is emailed a deepfake of themselves.