CAG in the Real World: Latency, Cost, and Limits Compared to RAG (Part 2/2)
- llm
- retrieval-augmented-gen
- artificial-intelligence
- machine-learning
Hi, I am Jawher Ennaceur - AI & Software Engineer
AI & Software Engineer
I build AI systems that get used, not just demoed.
I'm an AI & Software Engineer specializing in LLM-powered systems: RAG pipelines, multi-agent workflows, and knowledge graphs, deployed for real business use cases. Based in Tunisia. I care about shipping things that actually work in production, not just proof-of-concepts.
Professional Experience
DELOITTE
Tunis, TunisiaFeb 2026 – Jun 2026
AI Engineer
Internship- Built an AI-powered insurance sales assistant from requirements analysis to deployment, helping agents progressively qualify customer needs, check product eligibility and generate personalized recommendations.
- Designed a multi-agent system with LangChain and LangGraph, coordinating five specialized agents through shared state, with persistent workflow state and automatic recovery via LangGraph checkpointing on PostgreSQL.
- Combined LLM-based reasoning with a deterministic business-rules engine that enforces eligibility and commercial constraints, so decisions follow explicit, reviewable rules.
- Built a Graph RAG pipeline on Neo4j that turns unstructured documents into a knowledge graph through OCR, semantic processing and LLM-based structured extraction.
- Designed prompt-driven agents that generate tailored product recommendations and sales arguments; in tests, they surfaced 60%+ of previously missed cross-selling opportunities.
- Benchmarked proprietary and open-source LLMs on cost, performance and latency to choose the models for the solution.
- Developed the FastAPI backend and React/TypeScript frontend, with JWT authentication, role-based access control, agency-level data isolation and role-specific Metabase dashboards.
CLASS QUIZ
Sousse, TunisiaAug 2025 – Feb 2026
AI Engineer
Part-time- Automated a manual exam workflow (printed exams, handwritten answers, photos sent via WhatsApp) with an end-to-end pipeline covering page identification, content extraction and grading.
- Designed a marker-based page-recovery system: corner markers encoding exam ID and page number are added to each PDF page, then detected in student photos to identify, crop and rectify pages from noisy backgrounds.
- Built an OCR and exercise-segmentation pipeline with an 8-category LLM classifier and a dedicated extraction prompt per type, structuring blank exams (content, answer key, marking scheme) and handwritten student responses into PostgreSQL.
- Implemented RAG-grounded automated grading that evaluates each answer against the original exercise, correct answer, marking scheme and curriculum knowledge from a vector database of educational books, giving teachers consistent feedback and reducing manual grading effort.
- Built an automated book-summarization system that extracts chapter structures from encrypted educational PDFs using OCR and LLMs, reducing manual content-preparation time by about 70%.
- Deployed FastAPI web applications with interactive content-validation tools, letting teachers review and correct extracted content across grade levels and subjects.
Jun 2025 – Jul 2025
AI Engineer
Internship- Designed a RAG-based educational chatbot that assists children while they solve exercises, using the current exercise, its correct answer, conversation history (PostgreSQL) and content retrieved from official Tunisian curriculum books in ChromaDB.
- Built a contextualization step that rewrites each child's query using conversation and exercise context before retrieval, with separate prompting strategies for child and parent users on the same pipeline.
- Added a modular linguistic-correction layer that normalizes spelling and grammar errors in children's input, reducing input errors by about 40% on the messages I tested.
- Implemented voice interaction as an interface layer (speech-to-text, chatbot, text-to-speech) without changing the core text pipeline.
- Converted Unity-oriented JSON exercise data into Markdown with Python scripts so it could be reused as AI context.
- Unified all components behind a modular FastAPI backend.
DATA SCALE
Tunis, TunisiaJul 2024 – Aug 2024
Machine Learning Engineer Intern
Internship- Automated client onboarding for a SaaS purchase-intent prediction platform for e-commerce, building a preprocessing pipeline that adapts each new client's raw data (shared and site-specific features) to the existing ML workflow.
- Implemented missing-value imputation (forward-fill, median), target encoding for categorical features, and automated data cleaning and transformation.
- Designed an automated transfer-learning pipeline for a LightGBM model, pre-trained on general data and fine-tuned per new client, replacing manual per-client retraining and making onboarding reusable.
- Developed a RAG-based website chatbot with Google Agent Builder, backed by a categorized knowledge base, and integrated a booking API so users can reserve services directly in the conversation.
- Created Flask APIs to serve trained ML models and automate backend data-processing services.
AI & Data Projects
RAG Chatbot Assistant
Ask questions about your own documents and get short, grounded answers. Files are chunked and embedded into a vector store, and a Gemini model answers using only the retrieved passages and the conversation so far.
- RAG
- LLM Applications
- LangChain
- FastAPI
- Vector Databases
- NLP
Product Advisor — LLM Assistant
This project combines a large language model with structured product data to build a conversational shopping assistant. It uses intent classification and entity extraction to separate recommendation requests, product comparisons and general questions. Fuzzy product matching resolves product names in comparisons, and prompt-based LLM responses turn the matched specifications into practical advice on price, RAM, storage, camera and battery.
- LLMs
- LangChain
- Prompt Engineering
- Natural Language Processing
- Intent Classification
- Fuzzy Matching
Genetic Algorithm — Defense Scheduling Optimization
This project uses a genetic algorithm to automate the scheduling of academic defenses. It reads students, supervisors and rooms from Excel files, builds conflict-free candidate schedules, and provides selection, crossover and mutation operators to improve them. A custom evaluation function penalizes resource conflicts and scores how compactly each professor's sessions are grouped across days and time slots.
- Genetic Algorithms
- Optimization
- Python
- Constraint Satisfaction
- Scheduling
- Data Processing
Hand Sign Classifier
The notebook loads a class-per-folder image dataset (133 training and 32 validation images), applies random augmentation, and trains a small classification head on top of a frozen MobileNetV2 for 20 epochs. It plots training and validation accuracy and loss, saves the model as an HDF5 file, and passes a webcam frame captured in Colab through the model to get a two-class softmax output. Validation accuracy reaches 1.0, but on a 32-image split that is augmented like the training data, so it is a sanity check and not a measure of real-world accuracy.
- Transfer Learning
- Image Classification
- Data Augmentation
- TensorFlow And Keras
- Computer Vision
Face Verification with a Siamese Network
Instead of classifying faces into fixed identities, this project learns a face embedding and compares two images by the distance between their embeddings, an approach designed to verify a face against a single reference photo without retraining a classifier. Built with TensorFlow and a custom training loop.
- Deep Learning
- Computer Vision
- Siamese Networks
- Data Augmentation
- TensorFlow
Billing Fraud Detection
The notebook merges client records with invoice history, cleans and encodes the data, and undersamples the majority class to a 3:1 ratio. It then trains and compares three models on a stratified 75/25 split. All three reach 100% accuracy, because the label column was left among the input features, so the notebook demonstrates the full pipeline but does not yet measure real detection performance.
- Fraud Detection
- Feature Engineering
- Imbalanced Data Handling
- Scikit-Learn
- Neural Networks
- Exploratory Data Analysis
My Skills
AI & Machine Learning
- Machine Learning
- Deep Learning
- Computer Vision
- NLP
- LLM
- RAG
- Graph RAG
- Agentic AI
- Multi-Agent Systems
- Prompt Engineering
- Transfer Learning
Data
- Data Science
- Data Analysis
- BI
- ETL
- SQL
- NoSQL
Frameworks & Libraries
- LangChain
- LangGraph
- Flask
- FastAPI
- PyTorch
- TensorFlow
- Scikit-learn
Cloud & DevOps
- Docker
- Azure
- GCP
- MLOps
Software Engineering
- REST APIs
- Web Development
- Algorithm Design & Analysis
- System Design
- Authentication & Authorization
Programming Languages
- Python
- Java
- JavaScript
- TypeScript
- C++
- C
Tools & Frameworks
LangChain
LangGraph
FastAPI
Docker
Pytorch
GCP
Azure
PostgreSQL
Neo4j
ChromaDB
Academic Journey
Education
Sept. 2023– July 2026
Software Engineering (National Engineering Diploma)
Institution: Higher Institute of Applied Sciences and Technology of Sousse
Sept. 2021– June 2023
Pre-Engineering Cycle – Mathematics, Physics and Computer Science (MPI)
Institution: Higher Institute of Applied Sciences and Technology of Sousse
2017 - 2021
Scientific Baccalaureate
Institution: Eljem high school
Certifications
Nov. 2025
Building LLM Applications With Prompt Engineering
Institution: NVIDIA
March 2025
Generative AI with Diffusion Models
Institution: NVIDIA
April 2023
Deep Learning Specialization
Institution: DeepLearning.AI
Nov. 2022
Machine Learning Specialization
Institution: Stanford University & DeepLearning.AI
Blog & Articles
CAG in the Real World: Architecture of a Preloaded KV Cache System (Part 1/2)
- genai
- python
- machine-learning
- llm
Engineering Agentic AI Systems: What MCP Really Changes for Agents (Part 3/3)
- ai-engineering
- software-engineering
- artificial-intelligence
- llm
- ai-agent
Engineering Agentic AI Systems: Multi-Agent Orchestration with LangGraph (Part 2/3)
- ai-agent
- software-engineering
- artificial-intelligence
- llm
- machine-learning
Engineering Agentic AI Systems: When LLMs Meet Deterministic Rules (Part 1/3)
- llm
- software-engineering
- ai-agent
- artificial-intelligence
- machine-learning
Community Involvement
NATEG ISSATso Club
- Event Manager:Planned and executed club events from start to finish. Coordinated teams for marketing, organization, and logistics to ensure smooth delivery.
- Board Member:Participated in decision-making processes. Helped set club goals and strategies.
- HR Member:Managed member relations. Resolved conflicts and supported team dynamics.
- AI Instructor:Designed and delivered AI training sessions. Helped over 60 participants grasp AI fundamentals and practical applications.
Tunisian Scout Association
- Wood Badge Holder:Achieved advanced scouting skills and demonstrated commitment to the movement.
- Group Leader:Led scout groups. Organized activities and mentored younger members.
- Trainer:Delivered training sessions to develop skills and build confidence in scouts.
- Team Builder:Promoted teamwork and community values within the association.
Contact Me!
Prefer email? jawherennaceur@gmail.com