Axiom: I Built a Tool That Shows You How the Human Brain Responds to Your Content

(And no, this shi isn't all AI generated, it's hand written.)

Content creation is still a black box, no matter how many analytics you have. You never truly know what your audience feels when they watch your content. And what kind of content is 'popping' at the given time, is still a mystery.

I wanted to know what was actually happening inside the viewer's brain.

So I built Axiom.

It all started when i came across a video on Meta's TribeV2 model (A Predictive Foundation Model Trained to Understand How the Human Brain Processes Complex Stimuli, which i'll explain in a bit, dw) and thats when i exactly knew what i was gonna do for the next few weeks.

Axiom is a neurological content analysis platform that uses real fMRI-based AI to predict brain activation across every moment of a short-form video, visualise it in 3D, and tell you exactly what to change and why.

The Problem With Content Analytics Today

Before we dig deep into Axiom, lets first understand the problem i was trying to solve.

Every platform gives you the same metrics. Views, watch time, completion rate, saves. These are downstream signals they tell you what happened after the content has been uploaded and consumed, and not really what happens in between, wheather if it will be a hit or miss.

A video can have a 70% completion rate and still be losing the viewer's brain 15 seconds in. The algorithm rewards the output. Nobody measures the input which is the brain itself.

That's what Axiom tries to change.

What Is TRIBE v2?

Axiom is powered by Meta's TRIBE v2 a multimodal AI model trained on real fMRI brain scan data. TRIBE v2 was built by Meta's research team to predict neurological responses to video content.

Here's what makes it unusual: it wasn't trained on engagement proxies. It was trained on actual brain scans where real people watched videos inside an MRI machine and researchers measured blood oxygenation changes across the cortex. TRIBE v2 learned to predict those responses from the video content alone, and the whole model itself is built upon 3 pretrained ai models, one for visual, one for audio and one for transcription/text - fused via a transformer architecture

The model outputs activation predictions across 20,484 vertices on the fsaverage5 brain surface a standard neuroimaging mesh used in academic research. One prediction per vertex, per 1.49 seconds of video. That's the hemodynamic resolution of fMRI.

The result is a dense, spatially structured map of predicted brain engagement for every moment of your content.

read the full paper here

How does Axiom work

(When you Upload a Reel)

1. Neurological Inference

The video is sent to a A100 or a RTX PRO 6000 BW GPU hosted on modal.com where TRIBE v2 model runs the inference.

The model processes the visual frames, audio, and speech simultaneously outputting a brain activation map across all 20,484 cortical vertices for every 1.49-second window of the video.

This takes roughly 3–5 minutes for a 60-second Reel.

2. Attention Timeseries

Axiom extracts an attention signal by combining activation from the Dorsal Attention Network (voluntary focused attention) and Ventral Attention Network (stimulus-driven attention capture) against suppression of the Default Mode Network (mind wandering).

Basically, 2 regions of the brain that fire up when your brain is GENUINELY paying attention to something (DAN + VAN) and the 3rd region that fires up when your brain is not paying attention (DMN).

This DAN/DMN anticorrelation is one of the most replicated findings in cognitive neuroscience. When DAN is up and DMN is down the brain is genuinely engaged.

(DAN + VAN GOOD, DMN BAD ; THEREFORE => (DAN+VAN)-DMN = SCORE GOOD)

The result is a second-by-second attention curve for the entire video, with:

• Hook score mean attention in the first 3 seconds
• Peak moment the timestamp of maximum neurological engagement
• Drop point where attention falls fastest
• High attention percentage what fraction of the video held sustained engagement

3. Live 3D Brain Visualisation

The brain activation data is serialised as a float16 binary array (converted from its original float32 format), resulting in overall 50% compression with negligible data loss for colour mapping purposes, which helped in significantly reducing transfer times and storage costs without compromising the visualisation quality. This data is served to the browser from Cloudflare R2.

In the browser, Three.js renders a fsaverage5 inflated brain surface mesh with 40,004 vertices. As the video plays, vertex colors update at 60fps interpolating between TRIBE v2 frames to produce smooth transitions.

Blue means low activation. Red means peak engagement. You can literally watch attention move across the cortex in real time as your video plays.

4. AI-Powered Edit Suggestions

This is where it gets specific.

Axiom extracts one video frame per brain timepoint using OpenCV, sends them to Qwen2.5-VL-72B via OpenRouter for visual description, and runs Whisper to get the transcript.

It then merges three data streams into a unified timeline:

for example:

This combined timeline is passed to DeepSeek V3 with a structured prompt that asks for timestamp-specific, data-grounded edit suggestions.

And it dosent just spit out generic edit suggestions. It gives you specific edit suggestions, such as:

> "Cut everything from 40s onwards attention drops from 0.84 to 0.19 in one second. The transcript shows a closing statement and the frame is a static talking head. End the video at 38s on the high-engagement moment instead."

Every suggestion cites the exact timestamp, the brain signal, what was on screen, and what was being said.

The Architecture

Axiom is fully serverless. Not a single byte touches a server disk during normal operation.

Browser
  ↓ multipart stream (busboy)
Node.js backend (Render)
  ↓ stream directly to
Cloudflare R2
  ↓ presigned video URL
Modal A100/RTX PRO 6000 GPU (on demand)
  ↓ TRIBE v2 inference
  ↓ brain.bin(for the brain visualisation) + report.json streamed back to
Cloudflare R2
  ↓ presigned URLs returned to
Browser
  ↓ fetches directly from R2 CDN

Cloudflare R2 is the backbone. Zero egress fees means serving video, brain binaries, and reports globally costs nothing in bandwidth. Presigned URLs expire automatically no auth layer needed on the media routes.

Modal spins up an RTX PRO 6000/A100 GPU only when a Reel is submitted. Idle time costs nothing. A 60-second Reel costs roughly $0.55 in GPU compute.

Server-Sent Events streams GPU logs from the Modal container back to the browser in real time via a callback architecture, Modal 'POSTs' log lines to a Node.js webhook, Node appends to the job and pushes via SSE (server sent events).

The Stack

Cost Per Reel

Honesty About Accuracy

TRIBE v2 predicts brain responses with an average R² of 0.54 across the whole cortex. Visual and auditory regions are more accurate (0.70–0.85). Default Mode Network regions are weaker (0.35–0.45).

Soo what i did was instead of relying on multiple factors which vary a lot in their accuracy, I took the simpler and the straight-forward approach, which worked wonders for me.(taking only the attention signals in account, DAN, VAN and DMN)

The attention signal is directionally accurate it reliably identifies high and low engagement moments but the absolute values should not be treated as ground truth. There's also an unavoidable hemodynamic lag of 4–6 seconds: i.e, the brain data reflects responses to content from several seconds earlier, due to the delay in blood circulation in our brain.

But, For content analysis purposes finding your strongest moments, identifying dead zones, understanding where you lost the viewer: this level of accuracy is meaningful and actionable. It's not neuroscience research, It's a signal that's significantly more grounded than watch time curves.

What's Next

• Whisper transcript integration for richer suggestion context
• Persistent storage with a database backend
• Multi-video comparison analyse multiple Reels and compare neurological profiles
• Creator dashboard with historical analysis
• Support for YouTube Shorts and TikTok

Try It

Axiom is currently in early access. If you create short-form content and want to understand what's actually happening in your viewer's brain DM me.

Content creation is still mostly guesswork. This is an attempt to change that.

live demo

Built by Yousuf · Powered by Meta TRIBE v2 · Running on Modal · Stored on Cloudflare R2