Medium

Ampatishan Sivalingam

In

Towards AI

by

Ampatishan Sivalingam

·

May 22

Forcing 16×16 Patches into 4096D Text Manifolds: A Mechanistic Model of Multimodal Projection…

The quiet architectural crime at the heart of every vision language model you have shipped

Forcing 16×16 Patches into 4096D Text Manifolds: A Mechanistic Model of Multimodal Projection…

In

Towards AI

by

Ampatishan Sivalingam

·

May 21

Your Edge LLM is Memory Bound: Trading Compute for Bandwidth to Hit 30 Tokens per Second via LiteRT…

The Promise Nobody Warned You About

Your Edge LLM is Memory Bound: Trading Compute for Bandwidth to Hit 30 Tokens per Second via LiteRT…

In

Towards AI

by

Ampatishan Sivalingam

·

May 21

Tokenizing the Continuous: How Patch-Based Architectures Unlocked Zero-Shot Time Series at 200M…

Why the pointwise prediction era failed, and how a single architectural insight changed everything about how we forecast at scale

Tokenizing the Continuous: How Patch-Based Architectures Unlocked Zero-Shot Time Series at 200M…

In

Towards AI

by

Ampatishan Sivalingam

·

May 20

Rendering for Latent Space: Why Agent Native Browsers Swapped Pixels for Tokenized Accessibility…

How the browser engineering community stopped building for human eyes and started building for transformer weights

Rendering for Latent Space: Why Agent Native Browsers Swapped Pixels for Tokenized Accessibility…

In

Towards AI

by

Ampatishan Sivalingam

·

May 18

Benign Overparameterization: Deconstructing Why 100B Parameter Transformers Defy the Bias Variance…

The rule said bigger models must overfit. Then we built GPT and it forgot to read the rule.

Benign Overparameterization: Deconstructing Why 100B Parameter Transformers Defy the Bias Variance…

In

Towards AI

by

Ampatishan Sivalingam

·

May 18

Trading Zeros for Geometry: How Reshaping Transformer Weights to 2:4 Structured Sparsity Halves…

We spent the last five years chasing ghosts inside our weight matrices. And for a long time, we were convinced we were winning.

Trading Zeros for Geometry: How Reshaping Transformer Weights to 2:4 Structured Sparsity Halves…

In

Data Science Collective

by

Ampatishan Sivalingam

·

May 17

Evicting Trillion-Parameter APIs: The Inference Tradeoffs Driving Pinterest’s Pivot to Self-Hosted…

How the industry’s most expensive prototyping crutch became a production liability — and why small, sharp models running on your own…

Evicting Trillion-Parameter APIs: The Inference Tradeoffs Driving Pinterest’s Pivot to Self-Hosted…

In

Data Science Collective

by

Ampatishan Sivalingam

·

May 17

Test Time Compute Over Parameter Scaling: How Hermes Agent’s Self Correction Loop Captured…

The frontier just moved. And it did not move by adding more weights.

Test Time Compute Over Parameter Scaling: How Hermes Agent’s Self Correction Loop Captured…

In

Towards AI

by

Ampatishan Sivalingam

·

May 15

Dot Products as Routing Engines: Deconstructing Matrix Projections in O(N²) Transformer Attention

Understanding how transformers actually route information through massive sequences requires abandoning simplistic analogies and diving…

Dot Products as Routing Engines: Deconstructing Matrix Projections in O(N²) Transformer Attention

In

Towards AI

by

Ampatishan Sivalingam

·

May 10

Stop Flushing the KV Cache: How GitHub Trades VRAM for Compute to Cut Agentic Workflow Costs by 10x

The Era of Stateless Agents: Building Intelligence with Goldfish Memory

Stop Flushing the KV Cache: How GitHub Trades VRAM for Compute to Cut Agentic Workflow Costs by 10x

Ampatishan Sivalingam

Ampatishan Sivalingam

Machine Learning Scientist specializing in Computer vision and LLM. Bridging the gap between academic research and real-world application.

Following

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech