Looking for a job

Abdullah Kousa

@abdullahkousa

AI/ML engineer building LLM agents and evaluation pipelines in Python; created AutoAnalyst, an agent that runs data analysis from natural-language questions.

Syrian Arab Republic

Message

AutoAnalyst turns natural-language questions into executable Python and pandas analysis over CSV and Excel datasets. I designed its self-correcting execution loop and streamed results to a custom UI.

For SQLForge, I fine-tuned Qwen2.5-Coder-1.5B with 4-bit QLoRA on a single 8 GB GPU. SQL execution accuracy rose from 57.4% to 65.6%, and query crashes fell by 35%.

I also rebuilt the core of DoorDash’s delivery ETA model on its public dataset, reducing RMSE from 18.7 to 14.9 minutes with XGBoost. I added quantile regression to return calibrated delivery windows.

My projects span LLM agents, retrieval-augmented generation, forecasting, and computer vision, with open-source work published to GitHub, PyPI, and Hugging Face. I’m completing a B.Sc. in Informatics Engineering with an AI specialization at Arab International University.

Experience

Work history, roles, and key accomplishments

Other Projects

Various

Jan 2023 - Dec 2026 (3 years 11 months)

Developed multiple projects including Archivist-AI for natural-language photo search, a Smart Medical First-Aid Assistant, MedReport AI for RAG-based medical report analysis, a Diabetes Prediction model, and Air-Writing Recognition.

Python SigLIP Faiss EfficientNet fastAPI Whisper XGBoost SHAP Mediapipe LSTM

Text-to-SQL Fine-Tuned LLM

SQLForge

Jan 2026 - Present (10 months)

Fine-tuned Qwen2.5-Coder-1.5B with 4-bit QLoRA on Spider, raising SQL execution accuracy from 57.4% to 65.6% and cutting query crashes by 35%. Built an execution-based evaluation harness and self-correction loop, and shipped with Weights & Biases tracking and GitHub Actions CI.

Python QLoRA Qwen2.5 Coder PyTorch PEFT TRL Weights Biases fastAPI

Probabilistic Delivery-Time Model

DoorDash ETA Prediction

Jan 2026 - Present (10 months)

Rebuilt the core of DoorDash's production delivery-ETA model on their public dataset of 197k orders, cutting RMSE from 18.7 to 14.9 minutes. Engineered signals from DoorDash's engineering blog and added probabilistic forecasting with quantile regression for calibrated delivery windows.

Python XGBoost PyTorch scikit learn Pandas Gradio fastAPI Docker

AutoAnalyst Agent Developer

AutoAnalyst

Jan 2026 - Present (10 months)

Engineered an open-source agent that converts natural-language questions into executable Python/pandas code against arbitrary datasets, using an LLM tool-calling loop over a sandboxed interpreter. Designed a self-correcting execution loop that retries until valid output is produced, with streaming to a custom UI via SSE.

Python Groq API LLaMA 3.3 70B LLM Tool Calling Pandas Matplotlib fastAPI SSE PyPI

Real-Time License Plate Recognition

PlateVision

Jan 2025 - Present (1 year 10 months)

Engineered an end-to-end pipeline detecting traffic-violation classes with custom-trained YOLOv8, tracking vehicles with ByteTrack, and reading plates with EasyOCR. Improved OCR reliability through voting and fuzzy matching, and accelerated inference with CUDA FP16 and TensorRT.

Python YOLOv8 ByteTrack EasyOCR fastAPI CUDA FP16 TensorRT

ECG Cardiac Diagnosis with Explainability

CardioAI

Jan 2024 - Present (2 years 10 months)

Achieved 98.61% test accuracy across 4 cardiac classes by fine-tuning EfficientNetV2-S on 12-lead ECG images. Integrated Grad-CAM heatmap overlays and deployed as a Dockerized FastAPI service with a live Hugging Face demo.

Python EfficientNetV2 PyTorch Timm Grad CAM fastAPI Docker

Education

Degrees, certifications, and relevant coursework

Arab International University

Bachelor of Science, Informatics Engineering

2021 -

Pursuing a B.Sc. in Informatics Engineering with an AI specialization, expected to complete in 2026.

DeepLearning.AI

Build and Train an LLM with JAX, Large Language Models

Completed the course 'Build and Train an LLM with JAX', implementing automatic differentiation and multi-head attention from scratch in JAX.

Coursera

Deep Learning Engineering Specialization, Deep Learning

Completed the Deep Learning Engineering Specialization, covering TensorBoard diagnostics, custom PyTorch layers, mixed precision, GPU profiling, Vision Transformer fine-tuning, and learning rate schedules.