AI / MLOPS · CERTIFIED TRAINER

I LEARN BY BUILDINGML SYSTEMS, END TO END

Data science & AI, Al al-Bayt University. I like taking a project past the notebook, into an API, a container and a deployment, and learning from what breaks on the way.

See the work

Yahia Lababneh operating service robots at a technology conference — Yahia — Irbid, 2026

SELECTED WORK

THREE PROJECTS. THREE CONSTRAINTS I HAD TO WORK WITH.

Each one is laid out the same way: the constraint that shaped the problem, the decision it led to, and where it ended up.

OmniDiagexplainable clinical decision support

A clinical decision support system with two modules, coronary disease and dysglycaemia screening. Each answers refer, no referral or uncertain, with a calibrated interval and the reasons behind the answer.

Read the case study

WALKTHROUGH LIVE DEMO SOURCE

OmniDiag result panel reading Refer — confirmatory testing recommended, with the calibrated interval 85.9%–87.1% — the result panel: a decision and its calibrated interval, not a bare score

CONSTRAINT: No threshold and no risk bands. The probability's meaning does not transfer between hospitals, so the model answers with a decision, not a score.
DECISION: A spline GLM calibrated with Venn-Abers and a conformal decision split by sex. "Uncertain" counts as a referral everywhere, and about 40% of patients land there.
OUTCOME: AUC 0.802 coronary0.752 dysglycaemia

AI-Sourcing Hubcross-border B2B sourcing

Connects Jordanian importers to Chinese suppliers: a landed-cost engine, an OCR pipeline for supplier catalogues, and an LLM extraction chain with provider failover.

Read the case study

LIVE DEMO SOURCE

AI-Sourcing Hub landed-cost calculator: a product line, then the CIF stage and the customs-and-fees stage, line by line — cost breakdown — every line traceable to a tariff row

CONSTRAINT: A landed cost nobody trusts is a number nobody uses — it had to be auditable line by line.
DECISION: Three stages — CIF, customs, commercial — with AST-safe evaluation of the tariff formulas instead of eval.
OUTCOME: HS-code tariff tables; extraction survives a provider outage

EMG Gesture Classificationsignal processing under hard constraints

Four-gesture classification from a single forearm EMG channel with a small subject pool. Diagnose the data first, then clean it, then engineer features.

Read the case study

Bar chart: a classifier scores 45.0% on the real 20 Hz EMG recordings and 56.5% on synthetic Gaussian noise of the same shape; chance is 25% — real recordings scored below matched noise — the problem was the sampling rate

CONSTRAINT: One channel, small subject pool. No second sensor to fall back on.
DECISION: Test the data against matched noise before tuning models. That traced the failure to a 20 Hz sampling rate. After a firmware fix to 500 Hz: fifteen handcrafted features across five families.
OUTCOME: ≈55–60% on unseen subjects (LOSO)chance is 25%

HOW I CAN HELP

WHERE I FIT IN

FOR TEAMS HIRING

Someone who takes ML projects all the way to deployment, and learns from each one

I've built services and model pipelines end to end with FastAPI, Docker, PostgreSQL and MLflow, and documented the reasoning behind each decision.

certified trainer (TOT) ×2 · robotics trainer and invited speaker · three years in student teams · six months as team lead

See how I document a system

ALSO · FREELANCE

I also take on freelance projects for businesses: automation, document extraction and AI integration, built around real-world data.

Discuss a project

JOURNEY

HOW I GOT HERE

2025TOT certification ×2 · first place, Pickaxe
Jul 2025 — runningOmniDiag
Sep 2025 — Feb 2026Team Leader, Al3ahed Team
Jun 2026 — runningAI-Sourcing Hub