I build AI systems. Not because they are magical, but because someone has to clean the mess behind the magic.

• 🗣 Hamsa Speech Systems
  Designing and improving production-grade Speech & ASR pipelines.
  Accents, noise, silence, overlapping voices all the things datasets politely pretend do not exist.

Most of my work happens in places no demo video shows:

• Logs that explain nothing
• Metrics that improve while the model gets worse
• Experiments named final_v7_really_final

AI looks impressive from far away.
Up close, it’s mostly careful thinking and quiet panic.

• Speech Recognition & Audio Processing
• Arabic NLP Models (where the language is rich, ambiguous, and rarely forgiving)
• Computer Vision (Detection & Segmentation)
• Deep Learning Architectures
• Model Evaluation & Error Analysis
• Data-centric AI (where the real problems live)

I believe systems should leave traces behind them.

• Training logs (loss curves, WERs, sudden collapses)
• Model checkpoints saved five minutes before disaster
• Evaluation reports written with cautious optimism
• Visualizations that explain why, not just what

If a model has no artifacts,
it probably has no story or no truth.

Images are comforting.
They make progress feel real — even when it’s fragile.

I collect certificates the same way others collect bookmarks:

• As proof I once understood something
• And as a reminder that understanding fades if not practiced

Learning never ends.
It just changes its excuses.

All projects, experiments, and half-finished thoughts live here:
👉 https://ahpro7.github.io/Ahmed-Haytham/

Some projects solve problems.
Others explain why the problem was harder than expected.

Both are useful.

• Email: ahpro001@gmail.com

If I reply late,
I’m probably reading logs that feel personally offended by me.

Tools are loyal until the version changes.

In the end, the model converges. The loss decreases. The demo works.

And you still don’t trust it.

Which means you’re paying attention.