Shantanu Wani

Building reliable AI over flashy demos
AI researcher
Shipped ML systems end-to-end
Open to impactful ML roles

[Shantanu Wani @ShaantanuWani]

If a distance is Euclidean,
exp(−λ d²) becomes a valid positive-definite kernel.

Since dK isn’t Euclidean globally,
kernel tricks don’t automatically apply to all strings.

But for finite subsets? They do.

Kolmogorov complexity isn’t just information theory,
it has geometry.

[Shantanu Wani @ShaantanuWani]

What if “distance” between objects wasn’t geometric…

…but based on how hard it is to transform one into the other?

Marcus Hutter studies the geometry of Algorithmic Information Distance — and the result is surprising.

Is it Euclidean?

[Shantanu Wani @ShaantanuWani]

1. dK is not Euclidean globally.
You cannot embed all of (B*, dK) into ℓ2.
2. But any finite Euclidean point set can be scale-embedded into dK.
3. Many infinite-dimensional spaces cannot embed into it.

So dK is constrained, but expressive.

[Shantanu Wani @ShaantanuWani]

This paper proves something important:

Bigger embedding models aren’t inevitable.

With:
• Better initialization
• Direct embedding-space distillation
• Quantization-aware training

You can win on efficiency and quality.

That’s production-focused ML.

[Shantanu Wani @ShaantanuWani]

Everyone keeps scaling embedding models to billions of parameters.

Google DeepMind did the opposite.

They built a 308M parameter embedding model that:

• Tops MTEB multilingual
• Competes with models 2× its size
• Survives 4-bit quantization
• Runs efficiently on-device

[Shantanu Wani @ShaantanuWani]

Results (MTEB Multilingual v2):

• #1 under 500M parameters
• Competitive with models ~2× larger
• Strong even at 128-dim embeddings
• Maintains performance at 4-bit quantization

That’s rare for embedding models.

[Shantanu Wani @ShaantanuWani]

Full dataset accuracy:
GraphSAGE → 98.59%

With 24 missing nodes:
GraphSAGE drop → 8.8%
CNN drop → 29%
Bayesian drop → 39%

Key takeaway:

Inductive GNNs + local graph reconstruction > global deep models for infrastructure systems.

[Shantanu Wani @ShaantanuWani]

Most AI papers on power system faults stop at interval localization.

This one goes further:
It diagnoses the fault and pinpoints the exact fault location using GraphSAGE.

Here’s the core idea.

[Shantanu Wani @ShaantanuWani]

Using the entire graph for precise localization doesn’t work well.

So they rebuild a small subgraph using k-order adjacent nodes near the fault.

This:
· Reduces noise
· Forces local feature learning
· Improves regression accuracy

It’s structured inductive bias.

[Shantanu Wani @ShaantanuWani]

Most RAG systems don’t need re-rankers.

Before adding cross-encoders:

• Fix chunking
• Use hybrid search (BM25 + semantic)
• Improve embeddings
• Let your LLM filter

Add re-rankers only when precision actually matters (legal, medical, citations).

[Shantanu Wani @ShaantanuWani]

Read about a startup which was spending $15K/month on RAG rerankers.

They didn’t need them.

Here’s what most ML engineers misunderstand about cross-encoders and why it’s costing real money.

[Shantanu Wani @ShaantanuWani]

Stage 2 — Cross-encoder
• Query + document fed together
• Full transformer pass per document
• Better relevance scoring

Precision jumps (~60% → ~85%).

But here’s the tradeoff:
You now run a full model 100 times per query.

That’s where your infra bill explodes.