owda.io

Islam Owda builds retrieval systems that can be measured.

AI/ML engineer in Gaza, Palestine. I take retrieval-augmented LLM applications the whole way — tuned search, an automated evaluation pass, live cost and latency monitoring, and a stack that comes up with one command.

See how it was built Download CV
Retrieval, live A query drawing its five nearest cases out of the corpus — the operation every answer on AI Cat Vet starts with.
Currently
Open to AI/ML engineering roles, remote or on-site
Studying
B.S. Computer Engineering, Islamic University of Gaza
Working language
English (B2 upper-intermediate), Arabic (native)

AI Cat Vet

Built solo as the LLM Zoomcamp capstone at DataTalksClub

A question-answering app for feline health that refuses to make things up. Every answer is grounded in a veterinary case dataset, scored by a second model, and logged so its cost, latency and relevance can be watched over time.

It ships as four services in one Docker Compose stack — the Streamlit app, PostgreSQL for conversation telemetry, Grafana for the dashboard, and Kestra running the evaluation workflow. The interesting work was not the chat interface; it was proving the retrieval was good enough to trust, and keeping proof coming in after deploy.

Read the repository

Held-out mean reciprocal rank, after tuning field weights

baseline 0.800
tuned 0.856

4

services containerized together

8

panels on the Grafana dashboard

  1. Retrieve

    minsearch over the case dataset, with field weights boosted and re-measured against hit rate and MRR until the right case came back first.

  2. Generate

    An OpenAI-API call behind a safety-focused prompt that keeps the answer inside the retrieved cases and defers to a vet where it should.

  3. Evaluate

    An LLM-as-a-judge pass that scores answer relevance and compares candidate generation models against each other, not against a vibe.

  4. Monitor

    Conversations, cost, latency and thumbs from users land in PostgreSQL and surface on an eight-panel Grafana dashboard.

  5. Deploy

    One docker compose up brings the whole stack online, with Kestra orchestrating the evaluation run.

Also built, also led

What I work with

Tools I have shipped something with, not tools I have read about.

Languages

  • Python
  • JavaScript (ES6+)
  • Java

LLM and retrieval

  • RAG architecture
  • Prompt engineering
  • LLM-as-a-judge evaluation
  • OpenAI API
  • minsearch
  • Hit rate and MRR

Machine learning

  • scikit-learn
  • TensorFlow
  • Keras
  • OpenCV

MLOps and infrastructure

  • MLflow
  • Kestra
  • Docker and Compose
  • Git and GitHub

Data and monitoring

  • PostgreSQL
  • Grafana

Interfaces

  • Streamlit
  • Django
  • HTML and CSS

Background

2028

B.S. Computer Engineering, Islamic University of Gaza

Expected. Third year: 96 of 156 credit hours done, CGPA 88.44 / 100.

2026

LLM Zoomcamp, DataTalksClub

Completed, with AI Cat Vet as the capstone.

2026

MLOps mentorship, Zomra

In progress. Model tracking and deployment with MLflow.

2022

High school diploma

96.4%.

Training along the way

  • Probabilistic machine learning and its mathematics
  • Computer vision and neural networks, Gaza Sky Geeks — CNNs with Keras
  • Deep learning with Keras, Coursera (IBM Skills Network)
  • Machine learning deployment, Coursera
  • Essentials of machine learning, Gaza Sky Geeks — 60 hours
  • IT foundations with Python, Gaza Sky Geeks — 50 hours
  • Django web development, Vision Plus — 40 hours
  • JavaScript fundamentals
  • ICDL, Islamic University of Gaza — 40 hours
  • Duolingo English Test: 125

Have something that needs to be grounded, evaluated and shipped?

I read everything. A repository link or a rough problem is plenty to start with.

contact@owda.io