diff --git a/README.md b/README.md
index d2fb5e810436d..83825ae4c7f19 100644
--- a/README.md
+++ b/README.md
@@ -1,20 +1,20 @@
-
 # Academic Pages
 
 ![pages-build-deployment](https://github.com/academicpages/academicpages.github.io/actions/workflows/pages/pages-build-deployment/badge.svg)
 
 Academic Pages is a Github Pages template for academic websites.
 
+`jekyll serve -l -H localhost` 
 
 # Getting Started
 
 1. Register a GitHub account if you don't have one and confirm your e-mail (required!)
-1. Click the "Use this template" button in the top right.
-1. On the "New repository" page, enter your repository name as "[your GitHub username].github.io", which will also be your website's URL.
-1. Set site-wide configuration and add your content.
-1. Upload any files (like PDFs, .zip files, etc.) to the `files/` directory. They will appear at https://[your GitHub username].github.io/files/example.pdf.  
-1. Check status by going to the repository settings, in the "GitHub pages" section
-1. (Optional) Use the Jupyter notebooks or python scripts in the `markdown_generator` folder to generate markdown files for publications and talks from a TSV file.
+2. Click the "Use this template" button in the top right.
+3. On the "New repository" page, enter your repository name as "[your GitHub username].github.io", which will also be your website's URL.
+4. Set site-wide configuration and add your content.
+5. Upload any files (like PDFs, .zip files, etc.) to the `files/` directory. They will appear at https://[your GitHub username].github.io/files/example.pdf.
+6. Check status by going to the repository settings, in the "GitHub pages" section
+7. (Optional) Use the Jupyter notebooks or python scripts in the `markdown_generator` folder to generate markdown files for publications and talks from a TSV file.
 
 See more info at https://academicpages.github.io/
 
@@ -23,13 +23,13 @@ See more info at https://academicpages.github.io/
 When you are initially working your website, it is very useful to be able to preview the changes locally before pushing them to GitHub. To work locally you will need to:
 
 1. Clone the repository and made updates as detailed above.
-1. Make sure you have ruby-dev, bundler, and nodejs installed: `sudo apt install ruby-dev ruby-bundler nodejs`
-1. Run `bundle install` to install ruby dependencies. If you get errors, delete Gemfile.lock and try again.
-1. Run `jekyll serve -l -H localhost` to generate the HTML and serve it from `localhost:4000` the local server will automatically rebuild and refresh the pages on change.
+2. Make sure you have ruby-dev, bundler, and nodejs installed: `sudo apt install ruby-dev ruby-bundler nodejs`
+3. Run `bundle install` to install ruby dependencies. If you get errors, delete Gemfile.lock and try again.
+4. Run `jekyll serve -l -H localhost` to generate the HTML and serve it from `localhost:4000` the local server will automatically rebuild and refresh the pages on change.
 
 If you are running on Linux it may be necessary to install some additional dependencies prior to being able to run locally: `sudo apt install build-essential gcc make`
 
-# Maintenance 
+# Maintenance
 
 Bug reports and feature requests to the template  should be [submitted via GitHub](https://github.com/academicpages/academicpages.github.io/issues/new/choose). For questions concerning how to style the template, please feel free to start a [new discussion on GitHub](https://github.com/academicpages/academicpages.github.io/discussions).
 
diff --git a/_pages/cv.md b/_pages/cv.md
index 59bd03bdfb836..b4ea39a98d0d7 100644
--- a/_pages/cv.md
+++ b/_pages/cv.md
@@ -129,6 +129,6 @@ Skills
 Publications
 ============
 
-<ul>{% for post in site.publications reversed %}
+<ul>{% for post in site.publications %}
     {% include archive-single-cv.html %}
   {% endfor %}</ul>
diff --git a/_pages/portfolio.html b/_pages/portfolio.html
index d5c93d2e408c9..0f6684b5131bd 100644
--- a/_pages/portfolio.html
+++ b/_pages/portfolio.html
@@ -8,7 +8,7 @@
 {% include base_path %}
 
 
-{% for post in site.portfolio %}
+{% for post in site.portfolio reversed %}
   {% include archive-single.html %}
 {% endfor %}
 
diff --git a/_portfolio/2021-06-01-ner.md b/_portfolio/2021-06-01-ner.md
index 71b1cf05f355c..900a8cc3a9fb5 100644
--- a/_portfolio/2021-06-01-ner.md
+++ b/_portfolio/2021-06-01-ner.md
@@ -3,8 +3,71 @@ title: "NER for commands extraction"
 excerpt: "Apr 2021 - Jun 2021: NER for commands extraction <br/><img src='/images/ner_model.png'>"
 collection: portfolio
 ---
-Implementation of a deeplearning model to extract commands from a user's query
-Use of ktrain framework and fine-tuning of BiLSTM-Bert model to obtain a language processing model capable of extracting commands
 
+**Project Title: Implementation of a Deep Learning Model for Command Extraction from User Queries**
 
-Writing and submitting a paper to the MLPC 2021 challenge organized by James at ENSPY
+**Project Overview:**
+
+This project focuses on developing an advanced natural language processing (NLP) model capable of extracting actionable commands from user queries. By leveraging the capabilities of deep learning, we aim to build a robust system that can interpret and process human language with a high degree of accuracy.
+
+**Key Components:**
+
+1. **Model Selection and Framework:**
+   * We utilize the ktrain framework, which is a lightweight wrapper for Keras, simplifying the process of building, training, and deploying deep learning models.
+   * Our primary model is a BiLSTM (Bidirectional Long Short-Term Memory) combined with a BERT (Bidirectional Encoder Representations from Transformers) model. This hybrid approach takes advantage of BERT's contextual language understanding and BiLSTM's sequential processing capabilities.
+2. **Model Fine-Tuning:**
+   * The BiLSTM-BERT model undergoes fine-tuning to adapt to our specific task of command extraction. Fine-tuning involves training the pre-trained BERT model on our dataset to improve its performance in recognizing and extracting commands from user queries.
+   * This process includes adjusting hyperparameters, optimizing the learning rate, and incorporating dropout techniques to prevent overfitting.
+3. **Dataset Preparation:**
+   * We compile a comprehensive dataset comprising various user queries, each annotated with the corresponding commands. This dataset is essential for training and evaluating the model.
+   * Data augmentation techniques are applied to enhance the diversity and size of the dataset, ensuring the model generalizes well to new, unseen queries.
+4. **Model Training and Evaluation:**
+   * The training process involves multiple iterations, with regular evaluation using metrics such as precision, recall, F1-score, and accuracy to measure the model's performance.
+   * We implement cross-validation techniques to validate the model's effectiveness and ensure it performs consistently across different subsets of the data.
+5. **Deployment and Application:**
+   * Once trained, the model is deployed as part of a larger NLP system capable of interacting with users and executing commands based on their queries.
+   * The deployment phase includes integrating the model into a user interface, allowing for real-time command extraction and execution.
+
+**Outcome and Future Work:**
+
+* **Research Paper Submission:**
+
+  * The findings and methodology of this project have been documented and submitted as a research paper to the MLPC 2021 challenge organized by James at ENSPY (École Nationale Supérieure Polytechnique de Yaoundé).
+  * The paper detail the model architecture, training process, evaluation metrics, and results, providing valuable insights into the application of deep learning in command extraction tasks.
+* **Potential Enhancements:**
+
+  * Future improvements may include expanding the dataset, exploring other model architectures, and incorporating additional NLP techniques to further enhance the accuracy and robustness of the system.
+  * Collaboration with other researchers and participation in related challenges can provide opportunities for continuous learning and innovation in this domain.
+
+  **Project Title: Implementation of a Deep Learning Model for Command Extraction from User Queries**
+
+  **Project Overview:**
+
+  This project focuses on developing an advanced natural language processing (NLP) model capable of extracting actionable commands from user queries. By leveraging the capabilities of deep learning, we aim to build a robust system that can interpret and process human language with a high degree of accuracy.
+
+  **Key Components:**
+
+  1. **Model Selection and Framework:**
+     * We utilize the ktrain framework, which is a lightweight wrapper for Keras, simplifying the process of building, training, and deploying deep learning models.
+     * Our primary model is a BiLSTM (Bidirectional Long Short-Term Memory) combined with a BERT (Bidirectional Encoder Representations from Transformers) model. This hybrid approach takes advantage of BERT's contextual language understanding and BiLSTM's sequential processing capabilities.
+  2. **Model Fine-Tuning:**
+     * The BiLSTM-BERT model undergoes fine-tuning to adapt to our specific task of command extraction. Fine-tuning involves training the pre-trained BERT model on our dataset to improve its performance in recognizing and extracting commands from user queries.
+     * This process includes adjusting hyperparameters, optimizing the learning rate, and incorporating dropout techniques to prevent overfitting.
+  3. **Dataset Preparation:**
+     * We compile a comprehensive dataset comprising various user queries, each annotated with the corresponding commands. This dataset is essential for training and evaluating the model.
+     * Data augmentation techniques are applied to enhance the diversity and size of the dataset, ensuring the model generalizes well to new, unseen queries.
+  4. **Model Training and Evaluation:**
+     * The training process involves multiple iterations, with regular evaluation using metrics such as precision, recall, F1-score, and accuracy to measure the model's performance.
+     * We implement cross-validation techniques to validate the model's effectiveness and ensure it performs consistently across different subsets of the data.
+  5. **Deployment and Application:**
+     * Once trained, the model is deployed as part of a larger NLP system capable of interacting with users and executing commands based on their queries.
+     * The deployment phase includes integrating the model into a user interface, allowing for real-time command extraction and execution.
+
+  **Outcome and Future Work:**
+
+  * **Research Paper Submission:**
+    * The findings and methodology of this project will be documented and submitted as a research paper to the MLPC 2021 challenge organized by James at ENSPY (École Nationale Supérieure Polytechnique de Yaoundé).
+    * The paper will detail the model architecture, training process, evaluation metrics, and results, providing valuable insights into the application of deep learning in command extraction tasks.
+  * **Potential Enhancements:**
+    * Future improvements may include expanding the dataset, exploring other model architectures, and incorporating additional NLP techniques to further enhance the accuracy and robustness of the system.
+    * Collaboration with other researchers and participation in related challenges can provide opportunities for continuous learning and innovation in this domain.
diff --git a/_portfolio/2021-09-01-workflow-optim.md b/_portfolio/2021-09-01-workflow-optim.md
index 8a76267d48a2b..00d186b3a6cea 100644
--- a/_portfolio/2021-09-01-workflow-optim.md
+++ b/_portfolio/2021-09-01-workflow-optim.md
@@ -3,10 +3,11 @@ title: "Constraints optimisation of resources uses by tasks in workflows"
 excerpt: "Feb 2021 - Sep 2021: Financial Data Generation <br/><img src='/images/scheduling.png'>"
 collection: portfolio
 ---
+The organization and management of production is a prerequisite for the success of projects in the business and research world. In this process, the scheduling function aims to organize the use of resources to meet a demand or satisfy a prepared plan. Thus, some public and private companies use the scheduling function to cope with complexity, improve deadlines or even adapt to unexpected events. Although this problem has been widely addressed in the scientific and academic domain, its application in an industrial environment is very often minimalist. For example, complex constraints (resource limits and priority constraints) are not always taken into account. In this paper, we propose a solution that relies on adaptations of Boolean satisfiability problem-solving methods to encode the constraints and propose a task scheduling that minimizes the total execution time of a workflow. We were able to obtain a solution that takes into account the scheduling of workflows with constraints of resource limits, priorities, conditions and proposes a scheduling that makes the best use of resources in a relatively short time. The interest of this work is that reducing the execution time of the users’ tasks will allow a better management of the resources allocated in the infrastructures, an increase of the applications’ performances and by ricochet a great satisfaction of the customers. 
 
-Mathematical modeling of the concept of resources utilization for a resource management platform
-Optimization of a scheduler for fair resource allocation.
-Implementation of schedule for fair allocation of resources available.
+Keywords : task, workflow, scheduling, boolean satisfiability problem.
 
 Language: Python, Java, Google OR-Tools
 Task: Mathematical modelling of resources usages, Scheduler Implementation
+
+<img src='/images/scheduling.png'>
diff --git a/_portfolio/2022-09-01-cssl.md b/_portfolio/2022-09-01-cssl.md
index 7a224e0cbca16..900852e35fcd2 100644
--- a/_portfolio/2022-09-01-cssl.md
+++ b/_portfolio/2022-09-01-cssl.md
@@ -3,7 +3,7 @@ title: "Continual Self Supervised Learning through distillation and replay"
 excerpt: "Self-supervised learning provides strong visual representations from unlabeled data offline, but struggles in continual learning, which this study addresses using distillation, proofreading, and a prediction layer to prevent forgetting. <br/><img src='/images/cssl.png'>"
 collection: portfolio
 ---
-Without relying on human annotations, self-supervised learning aims to learn useful representations of input data. When trained offline with enormous amounts of unlabeled data, self-supervised models have been found to provide visual representations that are equivalent to or better than supervised models. However, in continual learning (CL) circumstances, when data is fed to the model sequentially, its efficacy is drastically diminished. Numerous ongoing learning techniques have recently been presented for a variety of computer vision problems. In this study, by utilizing distillation and proofreading, we tackle the extremely challenging problem of continuously learning a usable representation in which input data arrives sequentially. We can prevent severe forgetfulness and continue to train our models by adding a prediction layer that forces the current representations vectors to precisely match the frozen learned representations and an effective selection memory for proofreading previous data.
+Without relying on human annotations, self-supervised learning aims to learn useful representations of input data. When trained offline with enormous amounts of unlabeled data, self-supervised models have been found to provide visual representations that are equivalent to or better than supervised models. However, in continual learning (CL) circumstances, when data is fed to the model sequentially, its efficacy is drastically diminished. Numerous ongoing learning techniques have recently been presented for a variety of computer vision problems. In this study, by utilizing distillation and proofreading, we tackle the extremely challenging problem of continuously learning a usable represenhttps://zindi.africa/competitions/arm-unicef-disaster-vulnerability-challenge/datatation in which input data arrives sequentially. We can prevent severe forgetfulness and continue to train our models by adding a prediction layer that forces the current representations vectors to precisely match the frozen learned representations and an effective selection memory for proofreading previous data.
 
 Language: Python
 Task: Continual Leaning, Class-IL, Task-IL, Data-IL, Self-Supervised Learning
diff --git a/_portfolio/2023-01-01-semi-ssl.md b/_portfolio/2023-01-01-semi-ssl.md
new file mode 100644
index 0000000000000..2b84fc4f12ab2
--- /dev/null
+++ b/_portfolio/2023-01-01-semi-ssl.md
@@ -0,0 +1,61 @@
+---
+title: "Semi-Supervised Learning with few labels"
+excerpt: "Semi Self-Supervised Learning: improving the performance of self-supervised learning models, especially in scenarios where only a small amount of labeled data is available <br/><img src='/images/semi-ssl.png'>"
+collection: portfolio
+---
+
+**Project Title: Enhancing Performance of Self-Supervised Models with Limited Labeled Data**
+
+**Project Overview:**
+
+This project focuses on improving the performance of self-supervised learning models, especially in scenarios where only a small amount of labeled data is available. The core objective is to enhance the quality of the output representations from these models, thereby boosting their performance in various downstream tasks.
+
+**Key Objectives:**
+
+1. **Enhance Output Representations:**
+
+   - The primary goal is to improve the robustness and discriminative power of the features learned by self-supervised models, ensuring high-quality representations even with limited labeled data.
+2. **Combine Self-Supervision with Supervised Learning:**
+
+   - By integrating a small amount of labeled data into the training process, the project aims to synergize the strengths of both self-supervised and supervised learning paradigms, thereby enhancing overall model performance.
+
+**Methodology:**
+
+The project employs a strategic two-step approach to achieve these objectives:
+
+1. **Incorporate a Linear Supervision Layer:**
+
+   - **Initial Stage:** During the initial training phase, a linear layer is added on top of the self-supervised model. This layer leverages the available labeled data to provide supervised signals that guide the learning process. This step ensures that the learned representations are well-grounded in the labeled data, improving their quality and relevance to the target tasks.
+2. **Enforce High Correlation Among Same-Class Inputs:**
+
+   - **Correlation Enforcement:** To further refine the output representations, the model is trained to maximize the correlation between representations of inputs belonging to the same class. This objective encourages the model to produce similar representations for same-class inputs, thereby reinforcing its ability to distinguish between different classes effectively.
+
+**Tools and Technologies:**
+
+- **Python:** The primary programming language used for the implementation of the project.
+- **PyTorch:** The main deep learning framework utilized for building and training the models, known for its flexibility and robust support for dynamic computation graphs.
+- **PyTorch Lightning:** A higher-level interface for PyTorch that simplifies the training process and provides utilities for efficient experimentation, facilitating the implementation of complex training loops and model checkpoints.
+- **Weights & Biases (WandB):** A tool for experiment tracking and visualization, helping to monitor model performance, log experiments, and visualize the progression of training.
+- **Google Colab:** An online platform offering free GPU resources, which is used for running experiments and providing a collaborative environment for development and testing.
+
+**Domain Knowledge:**
+
+- **Computer Vision:** Understanding the principles and techniques used in image processing and analysis, which is crucial for designing effective self-supervised and semi-supervised learning models.
+- **Self-Supervised Learning:** Techniques where the model learns to represent data without labeled examples, typically by predicting parts of the input data from other parts.
+- **Semi-Supervised Learning:** Approaches that utilize a small amount of labeled data alongside a larger amount of unlabeled data to enhance learning efficiency and performance.
+
+**Project Report:**
+
+The project's results, methodologies, and findings are comprehensively documented in a report hosted on Weights & Biases. This report includes detailed explanations of the techniques used, experiments conducted, and the outcomes observed. The full report can be accessed [here](https://wandb.ai/alexneakameni/semi-supervised-learning/reports/Semi-Supervised-Learning--VmlldzoyNzU5OTg2?accessToken=ecq1xr8r68wrnq0rw9febl6kxhctox59d37jd4ree32qwq6kysprgqp63k8k7rk7).
+
+**Outcome and Future Work:**
+
+- **Enhanced Model Performance:**
+
+  - The project aims to push the boundaries of self-supervised learning by improving model performance in scenarios with limited labeled data, making the models more practical and effective for real-world applications.
+- **Potential Enhancements:**
+
+  - Future improvements may include exploring other model architectures, incorporating more sophisticated self-supervised learning techniques, and experimenting with larger and more diverse datasets to further enhance the robustness and generalizability of the models.
+  - Ongoing research and collaboration with other experts in the field can provide additional insights and opportunities for innovation in self-supervised and semi-supervised learning domains.
+
+<img src='/images/semi-ssl.png'>
diff --git a/_portfolio/2024-05-01-object-detection.md b/_portfolio/2024-05-01-object-detection.md
new file mode 100644
index 0000000000000..07789e33f62cf
--- /dev/null
+++ b/_portfolio/2024-05-01-object-detection.md
@@ -0,0 +1,78 @@
+---
+title: "Object Detection Using Transformers"
+excerpt: "Developing a Machine Learning Algorithm for Accurate Counting of Roof Types in Rural Malawi Using Aerial Imagery <br/><img src='/images/object-detr.png'>"
+collection: portfolio
+---
+
+---
+
+**Project Title: Developing a Machine Learning Algorithm for Accurate Counting of Roof Types in Rural Malawi Using Aerial Imagery**
+
+**Project Overview:**
+
+The people of Malawi have faced numerous natural disasters and climatic shocks in recent years, such as droughts, floods, and landslides. These events, compounded by the impacts of Covid-19 and other global issues, have severely affected the health and well-being of most Malawians. Rural areas, where more than 80% of the population resides, have been particularly hard-hit. Accurate mapping of flood extents and corresponding damages using satellite imagery has seen significant progress globally. However, there remain substantial gaps in accurately determining the number of affected populations, especially in rural Malawi. Traditional grass-thatched roofs in rural areas are often missed by algorithms using satellite or aerial imagery to count populations or identify buildings affected by floods.
+
+**Project Objectives:**
+
+1. **Accurate Counting of Roof Types:**
+
+   - Develop a machine-learning algorithm to accurately count the number of grass-thatch, tin, and other roofed houses in aerial (drone) imagery.
+   - Improve the estimates of affected populations during natural disasters, ensuring effective evacuation and aid distribution in rural Malawi.
+2. **Enhanced Disaster Response:**
+
+   - Provide accurate data that helps in improving response times and saving lives by ensuring that the most affected communities receive timely assistance.
+
+**Methodology:**
+
+The project employs several state-of-the-art machine learning models to achieve the objectives:
+
+1. **[Data Collection and Labeling](https://zindi.africa/competitions/arm-unicef-disaster-vulnerability-challenge/data):**
+
+   - Aerial imagery has been collected over parts of Malawi, and these images have been labeled to identify different types of roofs: grass-thatch, tin, and others.
+   - The dataset comprises 4,772 images for training and 2,045 images for testing.
+2. **Model Development and Training:**
+
+   - Several models are employed and evaluated for their performance in accurately identifying and counting different roof types:
+     - **DETR (Detection Transformer) 50**
+     - **CondDETR (Conditional DETR) 50**
+     - **DETA 50**
+     - **DETR 101**
+     - **Yolos (You Only Look One-level Small) Base and Small**
+     - **DETR-50-dc5**
+   - Each model is trained on the provided dataset, with specific attention to the diverse structures of rural houses in Malawi, which are typically circular or rectangular with mud-built walls and grass-thatched roofs.
+3. **Evaluation and Fine-Tuning:**
+
+   - Models are evaluated based on their accuracy in counting roof types, with specific metrics such as Intersection over Union (IoU) and other relevant performance indicators.
+   - Fine-tuning involves adjusting thresholds and improving model predictions to enhance accuracy, especially for minority classes like 'Other' roof types.
+
+**Tools and Technologies:**
+
+- **Python:** The primary programming language used for developing and implementing the models.
+- **PyTorch:** The main deep learning framework utilized for building and training the models.
+- **Google Colab:** An online platform providing GPU resources for running experiments and facilitating collaborative development.
+
+**Domain Knowledge:**
+
+- **Geospatial Analysis:** Understanding the principles of geospatial data and imagery analysis to effectively interpret aerial images.
+- **Machine Learning:** Expertise in developing and fine-tuning machine learning models for image recognition and object detection.
+- **Disaster Response and Management:** Knowledge of the requirements and challenges in disaster response to ensure the developed solution is practical and effective in real-world scenarios.
+
+**Project Report:**
+
+The project's results, methodologies, and findings are documented comprehensively, detailing the techniques used, experiments conducted, and outcomes observed. This documentation provides insights into the model's performance and highlights areas for future improvements.
+
+**Outcome and Future Work:**
+
+- **Improved Disaster Management:**
+
+  - By accurately counting roof types and estimating affected populations, the project aims to enhance disaster response strategies, ensuring timely and efficient aid distribution in rural Malawi.
+- **Model Generalization:**
+
+  - Future work includes improving the generalizability of the models to other regions and incorporating additional features to handle different types of disasters and dwelling structures.
+- **Continuous Improvement:**
+
+  - Ongoing research and collaboration with local authorities and disaster management agencies to refine the models and incorporate feedback for better performance.
+
+
+
+<img src='/images/object-detr.png'>
diff --git a/_portfolio/2024-07-01-specializing-llm-telecom.md b/_portfolio/2024-07-01-specializing-llm-telecom.md
new file mode 100644
index 0000000000000..ddb50769e9a9d
--- /dev/null
+++ b/_portfolio/2024-07-01-specializing-llm-telecom.md
@@ -0,0 +1,83 @@
+---
+title: "Specializing Large Language Models for Telecom Applications"
+excerpt: "Enhancing the Accuracy of Falcon 7.5B and Phi-2 on Telecom Knowledge Using the TeleQnA Dataset <br/><img src='/images/llm-telecom.png'>"
+collection: portfolio
+---
+**Project Title: Specializing Large Language Models for Telecom Applications**
+
+**Project Overview:**
+
+Large language models (LLMs) have made significant strides in text generation, comprehension, and interaction. While their integration into various sectors has been impressive, the telecommunications industry has seen limited adoption. This project aims to bridge that gap by enhancing the performance of LLMs on telecom-specific knowledge using the TeleQnA dataset. The competition focused on improving the accuracy of Falcon 7.5B and Phi-2 models in answering multiple-choice questions related to different telecom knowledge domains.
+
+**Project Objectives:**
+
+1. **Specialize Falcon 7.5B and Phi-2 on Telecom Knowledge:**
+
+   - Enhance the accuracy of Falcon 7.5B and Phi-2 models in answering telecom-related multiple-choice questions.
+   - Utilize methods such as Retrieval Augmented Generation (RAG) and prompt engineering to improve model performance.
+2. **Address Key Challenges:**
+
+   - Tackle the complexity and diversity of telecom-related questions.
+   - Mitigate LLM hallucinations and fabrications.
+   - Improve the explainability of LLM responses.
+
+**Methodology:**
+
+The project involved several key steps to achieve the objectives:
+
+1. **Data Understanding and Preparation:**
+
+   - Analyzed the TeleQnA dataset to understand the structure and content of the telecom knowledge domains.
+   - Preprocessed the dataset to ensure high-quality input for model training.
+2. **Model Specialization:**
+
+   - **Baseline Evaluation:**
+     - Evaluated the initial performance of Falcon 7.5B and Phi-2 models on the TeleQnA dataset.
+   - **Fine-Tuning:**
+     - Fine-tuned the models using the preprocessed dataset, focusing on telecom-specific knowledge.
+   - **Retrieval Augmented Generation (RAG):**
+     - Implemented RAG to enhance the models’ ability to provide accurate answers by retrieving relevant telecom documents.
+   - **Prompt Engineering:**
+     - Developed and tested various prompts to guide the models towards generating more accurate responses.
+3. **Mitigating Hallucinations:**
+
+   - **Training with Reinforcement Learning:**
+     - Used reinforcement learning techniques to penalize incorrect or fabricated responses during training.
+   - **Fact-Checking Mechanism:**
+     - Integrated a post-processing step to check the factual accuracy of the responses against a reliable telecom knowledge base.
+4. **Enhancing Explainability:**
+
+   - **Attention Visualization:**
+     - Implemented tools to visualize the attention mechanisms within the models.
+   - **Explainable AI Techniques:**
+     - Applied explainable AI techniques to understand the decision-making process of the models.
+
+**Tools and Technologies:**
+
+- **Python:** The primary programming language used for developing and implementing the models.
+- **PyTorch:** The main deep learning framework utilized for building and training the models.
+- **Google Colab:** An online platform providing GPU resources for running experiments and facilitating collaborative development.
+- **Hugging Face:** A platform for accessing and fine-tuning pre-trained language models like Falcon 7.5B and Phi-2.
+
+**Domain Knowledge:**
+
+- **Telecommunications:** Understanding the technical specifications and knowledge domains within the telecom industry.
+- **Machine Learning:** Expertise in developing and fine-tuning machine learning models for text recognition and language understanding.
+- **Explainable AI:** Knowledge of techniques to enhance the transparency and interpretability of AI models.
+
+**Project Report:**
+
+The project's results, methodologies, and findings are documented comprehensively, detailing the techniques used, experiments conducted, and outcomes observed. This documentation provides insights into the model's performance and highlights areas for future improvements.
+
+**Outcome and Future Work:**
+
+- **Improved Model Accuracy:**
+  - The specialized models showed significant improvements in answering telecom-related questions accurately, enhancing their utility in the telecom industry.
+- **Enhanced Telecom Applications:**
+  - The project paves the way for integrating LLMs into various telecom applications, improving customer support, network management, and other telecom services.
+- **Continuous Improvement:**
+  - Ongoing research includes refining the models further, expanding the dataset, and incorporating feedback from telecom experts to ensure practical and effective solutions.
+
+Code made avalaible [here](https://github.com/KameniAlexNea/object-detection-detr)
+
+<img src='/images/llm-telecom.png'>
diff --git a/images/llm-telecom.png b/images/llm-telecom.png
new file mode 100644
index 0000000000000..daee507661efd
Binary files /dev/null and b/images/llm-telecom.png differ
diff --git a/images/object-detr.png b/images/object-detr.png
new file mode 100644
index 0000000000000..fff6218e0944f
Binary files /dev/null and b/images/object-detr.png differ
diff --git a/images/semi-ssl.png b/images/semi-ssl.png
new file mode 100644
index 0000000000000..516032d747ec5
Binary files /dev/null and b/images/semi-ssl.png differ