The system

somehow, the system finds
its own mistakes.

Every recommendation logged. Every outcome checked. Every error corrected before it happens again.

The someboty workbench — Material Intelligence in action
How it learns

Predict. Activate. Observe. Adjust.

This is the Karpathy loop — the pattern behind every AI system that gets better with use. For someboty, it runs three times a week.

The Karpathy Loop — someboty material intelligence calibration diagram

the karpathy loop · predict → activate → observe → adjust · someboty · 2026

For someboty, it means: Scout makes a recommendation → Somerce activates a creator → the outcome is observed → the scoring weights adjust.

After a month: Scout isn't guessing. It's calibrated to your portfolio. After a year: it knows which creator types convert for which brands.

We call it somefink — the self-learning calibration engine that turns every activation into a better prediction.
Step 01 — Predict
Scout scores a creator
74 out of 100 for P.Louise. Scout has done this 148 times. Each time, it gets fractionally more calibrated to what Somerce actually needs.
Step 02 — Review
Libby sees the recommendation
Pre-acceptance. Not outreach. Not contact. A recommendation. Thirty seconds. Accept or reject. Her call, every time.
Step 03 — Observe
The outcome is measured
GMV generated. Return rate. Engagement. Brand safety. Scout's prediction is compared against what actually happened.
Step 04 — Adjust
Next prediction is better
The weights adjust. Not dramatically. Incrementally. The next creator Scout recommends for P.Louise is fractionally more aligned with how Libby actually thinks.
somefink

148 predictions. The learning has started.

The system has made 148 predictions about which creators fit which brands. It has already caught its own blind spots — a geography gate automatically rejected 100 creators who weren't UK-based before anyone needed to look. 48 qualified UK creators are loaded, scored, and waiting for their first activation.

The outcome data — which creators actually drove GMV, which ones returned product, which ones never posted — already lives inside TikTok Seller Center. The moment an agency connects that data, somefink starts learning from real results. Not after five weeks. From day one.

SOMEFINK SNAPSHOT — illustrative · week 1
@doveOtto: 79 · EchoTik: 60→ Pending
@charlottetilburyOtto: 73 · EchoTik: 54→ Brand flag
@sophiehannahOtto: 64 · EchoTik: 50→ Under-scored
@chloeburrowsOtto: 35 · EchoTik: 11→ Both missed
Predictions logged148
Auto-rejected (geo-gate)100
UK creators pending outcomes48
Illustrative dataset matched to Somerce workflow. someboty sorted.
Infrastructure

One scoring engine. Two applications.

The engine that qualifies creators for brand partnerships — weighing content style against audience fit, commerce history against compliance risk, brand safety against posting volume — is not a feature. It's infrastructure.

Joe Yates is building LiveHost — the operating system for live host scheduling and management. A marketplace where live commerce hosts and brands find each other.

The matching problem LiveHost solves — does this host fit this brand? — is structurally identical to the matching problem Scout already solves. Same scoring dimensions. Same calibration loop. Same feedback mechanism.

One infrastructure. Two applications. The intelligence transfers.
This section should feel architectural, not promotional. The observation is the persuasion.
What's next

The infrastructure is live. somefink is running.

Every creator Somerce activates is a new data point.

Every data point makes the next recommendation more accurate.

→ See the predictions live at /intelligence → See what Monday morning looks like at /somerce