SHOWCASE

Selected projects, in detail.

Here are selected projects presented in more detail: their context, technical choices and the parts of the implementation that mattered.

How this portfolio’s chatbot works

Edge AI · Android

OfflineLingo

View repository

QUESTION

How can emergency communication be supported when the network is unavailable and transcription uncertainty must not be hidden?

SYSTEM

An Android speech-to-text translation application running locally, designed for cross-border response scenarios.

CONTRIBUTION

Designed an end-to-end Android flow: voice capture, local transcription, translation and an interface that makes uncertainty understandable.

IMPLEMENTATION

whisper.cpp for on-device speech recognition.
llama.cpp with a quantized Qwen 2.5 model for local translation.
No Wi-Fi, mobile data or cloud API required to run.

RESULT

A standalone field-translation prototype for situations where a potential error must be noticed and verified.

DESIGN CHOICES

On-device transcription and translation, with no dependency on a cloud service.
Word-level and overall confidence indicators to make uncertainty visible.
High-contrast interface, one-record control, fast language switching and an emergency-vocabulary glossary.

KotlinAndroidSpeech-to-textOn-device AI

Multi-agent systems

Good News Agent

View repository

QUESTION

How can a personalised newsletter be assembled without reducing editorial selection to a simple LLM query and summary?

SYSTEM

A news-intelligence platform orchestrating search, source assessment, verification and newsletter composition.

CONTRIBUTION

Designed the agentic workflow and its safeguards, from document research through newsletter assembly for subscriber groups.

IMPLEMENTATION

A LangGraph workflow separating planning, collection, qualification, verification and writing.
FastAPI API and React interface around a SQLite database.
Deduplication rules and audit trails linking each selection to the sources consulted.

RESULT

A system that makes selection steps explicit instead of conflating text generation with editorial judgment.

DESIGN CHOICES

Planning research angles and parallel collection by sub-topic with LangGraph.
Explicit scoring for relevance, freshness, credibility, journalistic signal and constructive valence.
Fact checking, source deduplication, subscriber-group management and email delivery.

PythonFastAPILangGraphReactSQLite

Private AI · Android

Gemmory

View repository

QUESTION

What does a memory assistant look like when notes, answers and the model must remain on the phone?

SYSTEM

A local-first AI memory vault that turns raw notes into connected knowledge that can be queried with Gemma 4 on device.

CONTRIBUTION

Integrated Gemma 4 E2B with LiteRT-LM and built the memory, conversation and relationship-exploration features on Android.

IMPLEMENTATION

Incremental token streaming, multi-turn conversations and cancellable generation.
Interactive memory graph and backlinks for exploring relationships between notes.
Model size and SHA-256 verification, with local Room persistence.

RESULT

An application without a backend or cloud inference: notes, conversations and generated answers remain on the device.

DESIGN CHOICES

Local inference with LiteRT-LM, without a backend, account, cloud sync or fallback API.
Model installation checked by size and SHA-256, with the model kept off the main thread.
Conversations, backlinks and the knowledge graph persisted in Room and stored in the app’s private space.

KotlinAndroidGemma 4LiteRT-LMRoom

Robotics · Full stack

Robot Manager UI

View repository

QUESTION

How can a LeRobot environment be made practical for daily use, from hardware detection to episode recording?

SYSTEM

A web interface and API for LeRobot SO-101 / SO-100 arms, structured around real setup and experiment steps.

CONTRIBUTION

Designed a single operator flow, from hardware detection to building and reviewing robotic datasets.

IMPLEMENTATION

React/TypeScript frontend and FastAPI API, containerised with Docker.
Port, camera and calibration management before teleoperation.
Trajectory recording, replay and dataset administration in the same tool.

RESULT

An interface that brings together manipulations usually scattered across robotics scripts and commands.

DESIGN CHOICES

Guided management of USB ports and camera feeds before robot control.
Calibration, low-latency teleoperation and real-time visualisation in a tabbed interface.
A single pipeline to record, replay trajectories and administer datasets.

TypeScriptReactFastAPIDockerLeRobot

Medical AI

Ovarian Cancer Segmentation

View repository

QUESTION

How can tumour and metastasis regions be handled separately while accounting for a competition-specific medical metric?

SYSTEM

A segmentation pipeline for the PINKCC challenge, developed in a two-person team around training, evaluation and ensemble voting.

CONTRIBUTION

Developed U-Net segmentation models, validation pipelines and an ensemble strategy tailored to the two region types.

IMPLEMENTATION

One model specialised for each target region: tumour or metastasis.
Data preparation, augmentation, cross-validation, metrics and prediction visualisation.
Majority and weighted voting evaluated to combine model outputs.

RESULT

PINKCC Challenge finalist: 12th team out of 42, with a 0.5082 score on the Dilated Dice + F2 metric.

DESIGN CHOICES

A specialised model per region type, then combination through majority voting.
Preparation, augmentation, cross-validation, evaluation and prediction-visualisation pipelines.
Final score of 0.5082 on a metric combining Dilated Dice and F2.

PythonTensorFlowEnsemble learningMedical imaging

Applied AI · This portfolio

This portfolio & its AI assistant

View repository

The website you’re exploring is a Next.js application built with React and TypeScript: a bilingual portfolio, case studies, articles and interactive diagrams. I integrated an LLM assistant with document retrieval to make my work accessible through conversation.

THE QUESTION

How can visitors explore my work through a few questions, without invented experience or a paid inference budget?

MY APPROACH

A bilingual assistant that searches the public portfolio, answers with sources, and keeps a classic Q&A path available at all times.

How I built it

01
One content source
I organized experience, projects and articles as bilingual TypeScript content. The website and retrieval use that data, so portfolio updates also inform the answers.
Explore the code: One content source
02
Retrieval before generation
I normalize words, expand French and English synonyms, then rank relevant passages. Retrieval selects up to five passages; the server adds profile context when needed.
Explore the code: Retrieval before generation
03
An API that coordinates the answer
A Next.js route validates input with Zod, checks scope and reserves quota before calling OpenRouter. The prompt distinguishes experience, projects and articles, and requests numbered references.
Explore the code: An API that coordinates the answer
04
Streaming and a fallback
The browser receives server-sent events, with generation cancellation and source links. If the service fails or references are invalid, a prepared answer takes over when one is available.
Explore the code: Streaming and a fallback

Figure 01Data flow

From portfolio content to a sourced answer

Select an element to explore its role.

Read every explanation

Question: The browser sends the question and a short history to the server. Model credentials stay on the server; visitors never receive them.
Checks: The server validates the request and reserves an attempt: ceilings of 5 per visitor and 40 total per day. Redis shares counters across instances; without Redis, these limits apply separately to each instance.
Retrieval: Lexical retrieval selects relevant passages from my experience, projects and articles. Sources distinguish professional experience from topics I have written about. No model training or vector database is needed.
Generation: The model receives the question and selected passages. Only free models are allowed, with zero-price caps. Unavailability never triggers a paid-model fallback.
Answer: The answer appears progressively. Reference numbers are checked against the selected passages, and visitors can open the sources. Responses remain labelled as AI-generated.

01 / 05

Question

The browser sends the question and a short history to the server. Model credentials stay on the server; visitors never receive them.

When a check fails, predefined answers take over. A valid reference alone does not prove that every claim is correct.

THE IMPORTANT CHOICES

One content source for the website and retrieval, avoiding a duplicated biography that goes stale separately.
Atomic Redis counters, when configured, to share the budget across server instances.
A useful fallback: exhausted allowance, unavailable service or a failed check → classic questions without a model call.

Public source passages are sent to the provider only for AI answers. This site does not save conversations. Application checks limit abuse; DDoS mitigation also depends on the hosting platform.

Next.jsTypeScriptLexical RAGOpenRouterRedisEN / FR

ADDITIONAL PROJECTS

Selected projects, in detail.

OfflineLingo

Good News Agent

Gemmory

Robot Manager UI

Ovarian Cancer Segmentation

This portfolio & its AI assistant

How I built it

One content source

Retrieval before generation

An API that coordinates the answer

Streaming and a fallback

From portfolio content to a sourced answer

Local Chat

RAG Chatting UI

Credit Card OCR

Ship Image Classification

Ask about me