Aditya Sridhar

MLP in NumPy98.69% test accuracySmall CNN99.41% test accuracyCIFAR-10 CNN89.69% test accuracyResNet-56+9.8 pts from skip connectionsVAE-137.0 nats, test ELBODCGAN3.30 bits of 3.32 bits spreadLSTM forecaster1.06% error, an hour aheadMini-GPT2.11 bits per characterViT-Tiny82.68% test accuracyBlindAssist100% direction accuracyAmazon ML Challenge696 rank of 89,000+Offline voice assistant~1.3 s to the first wordProtein impact explorerBest Project Code for Life, VITSmart operator cab3 min overheat warningFraud detection0.954 PR-AUC, unseen monthsSIPCOV-2026 paperPresented, first authorICMLDE paperUnder revision, first authorPatents2 published, 7 more filedKornia3 mergedHaystack1 mergedtimm1 mergedfastembed2 in reviewOutlines1 in reviewDSPy1 in review

Aditya Sridhar · Machine learning

Models you can run,
not just read about.

I train neural networks from scratch and put them where you can push on them. The one behind this text is real: trained on MNIST, running in your browser now. Scroll to follow a forward pass, or draw it a digit.

Layer 1 · input · 784 pixels

Where it starts

Computer Science and Engineering at VIT Vellore, graduating 2027. Every network starts with a grid of numbers: 784 pixels. Right now this one is reading about me. Draw a digit below and it reads yours instead.

Waiting for a digit

  1. 0
  2. 1
  3. 2
  4. 3
  5. 4
  6. 5
  7. 6
  8. 7
  9. 8
  10. 9

About me →

Layer 2 · hidden · 128 neurons

A model zoo, all trained here

From an MLP written in NumPy to CNNs, ResNets and a small GPT. Each one is trained from scratch, every number names the run it came from, and each one runs in your browser.

What each of the 128 neurons learned to look for, from its 784 weights. Cyan pixels excite it, pink ones inhibit it. Tiles light up when their neuron fires for the digit passing through.

Open the zoo →

Layer 3 · hidden · 64 neurons

Projects, papers, patents

An offline navigation aid for visually impaired users, a first-author paper at SIPCOV-2026 and a second under revision, two published patent applications and seven more filed.

The same digit as each of the 64 neurons reads it. Cyan pixels pushed the neuron to fire, pink ones held it back; dim tiles were switched off.

Projects →Research →

Layer 4 · output · 10 classes

Fixes that shipped

Merged pull requests in Kornia, Hugging Face timm and deepset Haystack: the bug, the cause, the fix and the test that keeps it fixed.

  1. 0
  2. 1
  3. 2
  4. 3
  5. 4
  6. 5
  7. 6
  8. 7
  9. 8
  10. 9
Which pixels voted for each answer, and the probability it got.

Contributions →

Prediction

That was one forward pass.

784 → 128 → 64 → 10, trained in NumPy with backpropagation written by hand: 98.69% on the 10,000 MNIST test digits. The strongest connections are drawn. If you want to talk about the work, write to me.

How it was trained →Email ↗GitHub ↗LinkedIn ↗

Index

  1. L1ModelsEvery model trained from scratch and runnable in your browser.
  2. L2ProjectsSystems built end to end.
  3. L3ResearchPapers and patents.
  4. L4Open sourceMerged fixes in libraries people depend on.
  5. L5LearnIdeas taken apart step by step.
  6. L6AboutWho I am and how to reach me.