Meriem Melki.

Academic research comparing PANNs+BiGRU, EfficientNet-B0, and an Audio Spectrogram Transformer for 4-class respiratory sound classification on ICBHI 2017. Best AST config scored 0.6835, surpassing the published paper baseline of 0.6810.

• Deep Learning
• Audio Processing
• Transformers
• Python

A LangGraph DAG agent that turns a topic, YouTube URL, or uploaded audio/video into a translated Markdown blog post, with Whisper transcription. Built end-to-end with a FastAPI backend and React front-end.

• LangGraph
• Groq LLM
• FastAPI
• React
• Whisper

An agentic system combining Sysmon event collection with an ML layer to classify alerts, reduce false positives, and trigger automated responses, achieving 84.6% detection recall (11/13 MITRE ATT&CK techniques) with ~25s mean time-to-detect.

• LangGraph
• ML
• Python
• Sysmon
• SIGMA
• ELK

Fine-tuned a COCO-pretrained YOLOv8s-pose model via transfer learning (96.9% mAP@50) and benchmarked Random Forest, Gradient Boosting, and LSTM classifiers for 5-class behavior recognition.

• Deep Learning
• Computer Vision
• PyTorch
• YOLOv8
• scikit-learn

Distributed ETL pipeline in Apache Spark (SQL + Streaming) and Java processing San Francisco complaint data, with multi-line CSV parsing and batch/streaming analysis persisted to HDFS.

• Apache Spark
• Spark SQL
• Java
• HDFS
• Maven

Rigorously compared 5 architectures (Dense, CNN, LSTM, GRU, CNN+RNN hybrid) and 4 embedding strategies (GloVe, Word2Vec, FastText, TF-IDF) for sentiment classification. Best deep learning result: GRU + Word2Vec at 0.874 accuracy, though a TF-IDF + Logistic Regression baseline reached 0.90.

• TensorFlow
• Keras
• NLP
• Gensim
• Scikit-learn
• Python

A cloud-based IT Help Desk chatbot built with AWS Lex (generative AI), Lambda for slot validation and business logic, Cognito for authentication, and a Lex Web UI front-end. Optional multi-channel support via Twilio for SMS and WhatsApp.

• AWS Lex
• AWS Lambda
• Amazon Cognito
• Twilio
• Python

A production-ready car price prediction app for the Tunisian market, serving a tuned Extra Trees model through a Flask API with a React frontend for single and batch predictions.

• Python
• Flask
• Extra Trees
• Scikit-learn
• React

100 days of hands-on DevOps practice: Linux administration, Docker, Kubernetes, Ansible, Terraform, Jenkins CI/CD, and AWS — solutions and notes for each daily KodeKloud challenge, from SSH hardening to Terraform-managed CloudWatch alarms.

• Docker
• Kubernetes
• Ansible
• Terraform
• Jenkins
• AWS
• Linux

A collaborative full-stack medical appointment booking platform built with a 6-person team, featuring doctor ratings and admin/user notifications.

• Angular
• NestJS
• TypeScript
• SCSS