Rani Baghezza's profile picture

Rani Baghezza

AI Integration Engineer, PhD

rani@baghezza.ai

About Me

I am an AI engineer with a PhD in Deep Learning and cloud expertise.
I bring clarity, simplicity, and automation to your business workflows.

As a business grows, people, processes, and data sources multiply.
Many businesses turn to AI too quickly, as a way to accelerate their workflows.
The consequence: 90% of AI projects don't make it to production.

To me, the main issue is starting with AI.
I am a firm believer in starting with processes, data, and people instead.

My process is simple: I start by understanding your business, goals, and strategy.
After that, I understand your people, data, processes, and the tools you are using.
Once I have a clear picture of your business, I can help you identify the repetitive tasks that can be automated, and the data sources that can be exploited.

With a holistic approach, I make sure that any AI project is aligned with your business goals, and brings tangible value to your business.

View Portfolio (PDF)
rani@baghezza.ai
Schedule a Meeting

My Services

Featured Projects

AI Assistant in Microsoft Azure linked to 100GB of internal medical data

AI Assistant in Microsoft Azure linked to 100GB of internal medical data

Context

Velox Operations had acquired a company with over 20 years of historical market research data in the medical industry. I have worked together with their CEO and COO, as well as a business partner to build an internal AI assistant aimed at leveraging over 100GB of medical data to speed up the work of their market research team, while complying with HIPAA regulations.

Outcomes

Scalabe and reliable AI assistant deployed in Microsoft Azure, using Chat GPT-4o and Azure OpenAI, with a custom branding. Addition of a PDF drag and drop feature and RAG capabilities.

Client's Website

Technologies

React Python Azure OpenAI RAG
LLM+RAG system to automate invoice processing

LLM+RAG system to automate invoice processing

Context

I have worked with Nuxly to build a system that automates invoice processing. LlamaParse and OCR are used to extract entities from invoices of different formats, and a Pinecone vector database together with a retrieval system are used to match invoices to clients.

Outcomes

POC that turns a PDF and a base of clients into a PDF to a JSON directly exploitable in the Odoo ERP.

Client's Website

Technologies

LlamaParse OCR Pinecone LangChain Python OpenAI FastAPI Mistral pytesseract Odoo
Optimization and embedded deployment of a car detection model for bicycles

Optimization and embedded deployment of a car detection model for bicycles

Context

I have worked with Survue to improve the accuracy of the car detection model deployed on bicycles, aimed at cyclists' safety. I have re-written an implementation of MobileNetV2-SSD in PyTorch, and quantized it manually. I have also mentored an intern from Northeastern University to teach him the ropes of MLOPs.

Outcomes

Increased the accuracy of the model, and built a MLOps pipeline to automate the training and deployment of the model. Switched to a new stack.

Client's Website

Technologies

PyTorch Tensorflow MLOps ONNX Quantization Mentoring Bash
Improvement of an embedded vision worfklow - MLOPs pipeline

Improvement of an embedded vision worfklow - MLOPs pipeline

Context

I have worked with Genalyte to improve the vision workflow of their Merlin product. The former system was using an inefficient classification model, which I've replaced with a lightweight, quantized object detection model running on a Raspberry Pi, with robust training and invariance to channel flipping. The final solution was deployed in a constrained environment by only modifying a single python script in production. OCR was added to parse error codes in real-time.

Outcomes

Improvement of state and error detection to a finer granularity. Bug fixes from previous versions.

Client's Website

Technologies

Tensorflow TFLite Raspberry Pi Linux MLOps Python Bash
Real-time stock pattern detections using Computer Vision

Real-time stock pattern detections using Computer Vision

Context

I have worked a client to fine-tune a vision model to detect candlestick patterns on a stock market chart. Stock market data was retrieved from Yahoo Finance and Coinbase, plotted, and annotated before using it to train YoloV8. A real-time detection system was built, with the ability to re-train the model with new data over time.

Outcomes

Accurate stock pattern detection for simple patterns, such as head and shoulders, double and triple tops and bottoms.

Technologies

Yahoo Finance API Coinbase API YOLOV8 FastAPI MLOps Python
Real-time LLM-based quest generation in a Pygame RPG

Real-time LLM-based quest generation in a Pygame RPG

Context

I have built a basic RPG game in Pygame to experiment with dynamic quest generation using LLMS. The challenge was to generate quests that are coherent in the game universe, and that can be linked with existing game entities at runtime so the players can complete them and get rewards. I have experimented with OpenAI, and Groq for fast inference, as well as a Postgres database with a pgvector extension to store the vector embeddings of the game entities.

Outcomes

Real-time quest generation, within a coherent game universe.

Technologies

Pygame Groq Postgres pgvector embeddings RAG Python
Real-time mesh reconstruction in mixed reality using RGB-D data

Real-time mesh reconstruction in mixed reality using RGB-D data

Context

During my postdoc at Université de Sherbrooke, I have worked as a ML Researcher to build a real-time mesh reconstruction system in mixed reality using RGB-D data. To achieve this, I have used the HL2SS library to capture data from the Hololens 2, and a mix of multi-threaded object detection (YoloV8), segmentation (FastSAM), 3D clustering (DBSCAN), and mesh reconstruction (Poisson Surface Reconstruction). The application was an initiative from VMWARE, and aimed at assisting industry worker in manual tasks using Mixed Reality.

Outcomes

Real-time mesh reconstruction in mixed reality, with a focus on accuracy and speed.

Technologies

Python Unity Visual Studio HL2SS Object detection Segmentation 3D clustering Mesh reconstruction
Strategic AI Consulting

Strategic AI Consulting

Context

I have worked with multiple clients as a strategic AI consultant to help them integrate AI into their workflows. I have helped SOPREMA Canada cut through the uncertainty of a complex R&D project, and guided them in the migration of their chatbot ecosystem to Microsoft Azure. I have helped Bill App France think about ways to use computer vision in fast food order assembly, in order to minimize forgotten items, and thus, client complaints. I have done consulting calls for various clients and mentored students in their AI projects.

Outcomes

Strategic report, guidance, and feasibility study, mentoring.

Client's Website

Technologies

Strategic Consulting Mentoring
AI Agents in a Godot video game

AI Agents in a Godot video game

Context

I am currently experimenting with AI agents in a simulated universe built from scratch in Godot. Currently, the NPCs are using GOAP (Goal-Oriented Action Planning) to prioritize goals, such as eating, or working. I am working on extending the game to add a progression for the player, as well as emotional and social modules for the NPCs, relying on LLM calls to generate realistic behaviors.

Outcomes

Prototype of a game with intelligent NPCs

Technologies

Godot GOAP Game design Game development

Get in Touch

rani@baghezza.ai
Schedule a Meeting