Aashish Paliwal

Building Keline (AI Brand Governance), PhD Scholar at @iiit_hyderabad, Open Source

[Aashish Paliwal @aa_paliwal]

My boss said dress for job you want not the job you have. Now I'm sitting in a disciplinary hearing dressed as BATMAN

[Aashish Paliwal @aa_paliwal]

At Keline, we reduce with UMAP before clustering the approved set into modes.

Global variance is not the same thing as "which visual neighbourhood does this creative belong to?"

That difference is why PCA alone is usually the wrong default here.

[Aashish Paliwal @aa_paliwal]

If you need a lower-dimensional space for brand modes, why not just use PCA?

It is the default answer in a lot of ML stacks. Compress the embedding. Keep the top variance.

For brand visuals, stopping there is usually the wrong default.

[Aashish Paliwal @aa_paliwal]

Put them side by side for this job:

Use PCA if you care about a compact global summary of variance, or as a light denoise before something else.

Use UMAP if you need a local map of styles so you can find modes and score against the one an asset landed in.

If the approved set is tiny, neighbourhood methods get noisy. PCA (or no reduce) can be safer.

[Aashish Paliwal @aa_paliwal]

@chrisegg Clean and cheap is mostly solved. The pick step is where brand usually breaks. Polished is not the same as on-brand.

[Aashish Paliwal @aa_paliwal]

The obvious failure in brand embedding space is "far away".

The weird one is a nearest neighbor that looks close by cosine, then feels wrong once you look at which visual mode it actually sits with.

[Aashish Paliwal @aa_paliwal]

Same problem here. Building for it at my startup.

Approved work is the baseline. A human still decides when something unfamiliar is actually right. Once they override and ship it, that goes into the fingerprint.

Next time something similar shows up, you can check against that. Not start from zero.

[Aashish Paliwal @aa_paliwal]

@lolitataub Grabbed the Kindle edition. Looking forward to reading it this weekend. Thanks for putting this out.

[Aashish Paliwal @aa_paliwal]

@DataInfraxInc @hazhubble Yep. The score is the easy part. What actually matters is who overrode it and why, otherwise you just get a nicer looking black box.

[Aashish Paliwal @aa_paliwal]

@johniosifov Creation scaled. Review didn't. If governance stays a separate queue after generate, the bottleneck just moved downstream.

[Aashish Paliwal @aa_paliwal]

@martechismktg This is the part guidelines never survive. Adjectives don't constrain a generator. What you actually approved does.

[Aashish Paliwal @aa_paliwal]

@_devisha It's pretty good. Using it as a small ops desk: research, outreach drafts, calendar, invoices, etc. One coordinator agent only pulls me in for decisions. Cuts a lot of overhead.

Burns through limits pretty fast though.

[Aashish Paliwal @aa_paliwal]

At Keline, we start from SigLIP 2 embeddings, reduce, discover modes in the approved set, then score a new asset against the mode it belongs to.

High-dim distance still matters as a guardrail when something is strongly off-brand.

More dimensions aren't the goal. A neighborhood you can actually audit in is.

[Aashish Paliwal @aa_paliwal]

Brand consistency scoring sounds like a distance problem: how far is this image from the brand?

So you embed everything in a high-dimensional space and compare vectors.

Why isn't that enough on its own?

[Aashish Paliwal @aa_paliwal]

UMAP is one way to do that.

It preserves which assets sit near which other assets, so density-based clustering can find the modes without you picking K in advance.

Then the geometric check is distance to the nearest mode, not the mean of every reference in the brand.