Available for Applied AI rolesVienna · EU remote

Mohamed Ali / Applied AI Engineer

I take models apart to find out which part is doing the work.

Retrieval systems, sequence-model ablations, and vision models — built, measured, and reported with the uncertainty left in. The digit classifier on the right is one of mine, written from scratch in C++.

  • 4th / ~200 image-restoration competition, on an 8GB card
  • 12 × 5 xLSTM configurations × seeds, mean ± 95% CI
  • 9 months applied AI at Austria's central bank
Live forward pass · 784 → 100 → 10my C++ weights, in your browser
1 · raw 28²
2 · fit 20²
3 · mass-centred
0
0.98 activation · sample labelled 0
00.9810.0020.0530.0040.0050.0360.0070.0080.0090.06

MNIST digits were size-normalised into a 20×20 box and shifted so their centre of mass sits at the centre of the 28×28 field. Feed a raw stroke straight in (1) and the network reads the wrong distribution, so the drawing is cropped and rescaled (2), then translated by centre of mass (3). Step 3 is what the network actually sees.

or load a sample
PROJECTS / 01—06

Selected projects

Applied systems and controlled experiments, with the implementation and evidence kept visible.

01Applied AI · Retrieval · Evaluation

Built a document-grounded RAG pipeline at OeNB.

A specialized chatbot workflow that grounds answers in an OeNB document collection and measures whether retrieval finds the right evidence.

Open case study
02Sequence models · Experimental research

Isolated what xLSTM blocks contribute.

A capacity-matched study of two-block xLSTM, LSTM, and Transformer models on associative recall and formal-language tasks.

Open case study
03Computer vision · Image restoration

Placed 4th in an image-restoration competition with roughly 200 students.

A from-scratch image-restoration entry that placed fourth while training on an 8GB GPU.

Open case study
04Audio ML · Feature fusion

Compared bird-call models under class imbalance.

An experimental comparison of model families, hyperparameters, acoustic representations, and evaluation criteria for bird-species recognition.

Open case study
05Visual analytics · Geographic data

Mapped new-coder geography across 91 countries.

Two country maps compare 2018 HDI with the percentage of surveyed new coders living in cities of more than one million people.

Open case study
06C++ · Neural-network fundamentals

Built a neural network from scratch in C++.

A handwritten-digit neural network implemented from first principles in C++ as a hands-on project for fun.

Open case study
PROFILE / CURRENT

From document retrieval to model internals.

At OeNB, I worked on document processing, retrieval, language-model workflows, and retrieval evaluation for RAG systems.

At JKU, I completed a Bachelor's in Artificial Intelligence and am continuing with a part-time Master's while building deeper experimental work.

Read the full profile
AVAILABLE / VIENNA + EU REMOTE

Looking for someone who can build the system and explain why it works?

Start a conversation