Aashish Paliwal (@aa_paliwal) profile picture

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

India
Joined June 2010
Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by Aashish Paliwal (@aa_paliwal)Photo by 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
10
114
4
186
Once UMAP gives you a small space where brand styles stay local, the next question is simple: how many styles does this brand have? K-Means wants that number up front. For brands, you usually don't know it.
1
1
80
One lesson from shipping: the minimum cluster size has to scale with how many references a brand has. Keep it fixed and small, and a larger reference set fractures into lots of tiny clusters that aren't real styles.
1
1
2
At Keline, HDBSCAN reads modes in the UMAP space, and we score a new asset against the mode it lands in. Which leaves the obvious next question. If something lands in noise, is it off-brand? Or just new?
1
2
SyPy group from @iiit_hyderabad is coming to Devcon 8. I'm here early. If you're working in security or privacy, come say hi. Know more about the research group at sypy.iiit.ac.in
1
6
249
Your API accepting 50 requests at once does not mean the GPU can run 50 inferences at once. HTTP concurrency is open connections. GPU concurrency is how many models actually fit and stay stable. Treat those as the same number and the queue looks healthy until inference starts falling over.
1
67
One reference image can't define a brand. It tells you what one asset looks like. It doesn't tell you which parts are on purpose and which are just that day's shoot. You need enough approved work to see what repeats. That's what a brand actually is.
78
When people hear "dimensionality reduction," they often picture a 2D scatter plot for a slide. Useful for humans. Not why it sits in a brand audit pipeline. So what is UMAP doing if it isn't for visualization?
1
93
Once you treat UMAP as representation, a few production rules show up. Fix the seed so the same refs and the same test image do not drift across restarts. And do not trust the reduced space alone for strongly off-brand work. High-dim distance still acts as a guardrail.
1
7
At Keline, SigLIP 2 embeddings still carry the visual detail. UMAP turns that into a space where modes are readable enough to cluster and score against. The plot is optional. The representation is not.
9
Putting ML inference inside the same request path as your API looks clean on day one. Then concurrency, retries, and long-running models start fighting the HTTP lifecycle. The painful part isn't loading the model. It's pretending a request and an inference job are the same thing.
1
119
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.
1
95
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.
1
49
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.
53