My projects :


Log project

This project implements a robust and centralized logging system designed to monitor, trace, and diagnose the behavior of a FastAPI-based backend application.

The goal is to provide clear visibility into both runtime operations and error events while maintaining clean separation between business logic and infrastructure-level concerns."

  • Frontend: React TS, tailwind, axios
  • Backend: Fast API, Pydantic and Websocket
  • DB: Opensearch for vector store
  • Containerization: Docker
Repo Github

Never-Answer

"LLM users performed worse at every level: neural, lunguistic and scoring."

That's MIT Media LAB, 2025:

"Your brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task"

And human brain grows through effort - not answers

What if students could harness AI's power while strengthening - not surrendering - their minds ?

Never Answer doesn't give students the answers, it gives then the right kind of difficulty.

  • Frontend: Next.js
  • Backend and AI API: Fast API
  • DB: FAISS for vector store
  • Containerization: Docker
Repo Github

Research: Neuronnal Pathway Coverage


Neuronal Pathway Coverage is a research project carried out as part of a hackathon organized by ApartResearch in the field of AI Safety.

Abstract:

This paper explores the mechanistic interpretability of large language models (LLMs) by analyzing the classification of political tweets. Using Llama 3.3 70B, we infer users' political alignment (Pro-Biden, Pro-Trump or Neutral) from their tweets, descriptions and locations. We then apply the Goodfire framework to extract the most activated features corresponding to each class. By re-running the model on the same dataset while activating these extracted features, we analyze the influence of these neural activations on classification. Our approach provides a better understanding of the paths taken by the LLM during classification, shedding light on the internal mechanisms linked to political biases in the model's predictions. We validate our results using accuracy comparisons and neural path coverage graphs.

  • Langage: Python
  • API: Goodfire
Repo Github

LaijaAI

LaijaAI is an assistant that generates a character.

Give her a backstory, a character, and she'll interpret it perfectly!

  • Frontend: React
  • Backend: Django
  • API AI: FastAPI
  • DB:
  • Containerization: Docker
Repo Github

Pong Online

Pong Online is a website for online pong games.

It also features instant messaging.

  • Frontend: React
  • Backend: NestJS
  • DB: Postgres
  • Containerization: Docker Compose
  • Authentication: 2FA

For pong and instant messaging, the protocol used is websocket.

The rest is REST API.

Repo Github

See my other project: Repo Github