AI engineer
Agentic systems

DocHub converts PowerPoint, PDF and Office files to Markdown, keeping images and attachments together. I developed it at GYS so teams could prepare their documentation without manually reworking every file. The desktop application lets them submit documents, compare the results and prepare publication to GitLab. AI restructuring is optional. I separated the interface, converters and publication modules so an engine can be replaced or a destination added without rebuilding the application.

Delivered at GYS; conversion choices studied across 981 slides from 57 presentations.

When a courier runs out of cargo, could a transfer from another courier avoid a trip back to the depot? I developed this personal project to investigate that question using synthetic delivery scenarios. The tool assigns customers, calculates routes under operational constraints and makes the results explorable on a map. My contribution combines transfer modelling, comparisons against external reference problems and experiments on how much solution quality improves with additional computation time. The replay connects those measurements to individual routes.

Solomon reference comparisons, controlled search budgets and interactive route replay.

View the demo Read the full story

An accessory disconnected during an update can keep an older software version and stop working correctly with the welding station. At GYS, I developed startup software for a board that stores a copy of the updates inside the equipment and sends them to the connected devices that need them. I integrated this behavior with the existing update system, working with the design office. A USB drive still takes priority when it supplies a new update.

Integrated on a prototype, with locally stored updates and USB update priority preserved.

An agent that can open a page still needs to find the right elements, act and understand the result. For my work with Hermès, I extended the Camoufox MCP adapter with a layer that interprets instructions such as “read the prices”. My contribution covers semantic tools, model routing, prompts and navigation diagnostics. The browser and its anti-detection capabilities come from the upstream projects, which remain the foundation of the service. The added tools help agents decide how to continue a changing workflow.

Fork of redf0x1/camofox-mcp, with semantic tools, prompts and diagnostics available in the public repository.

Read the full story

A listing price does not tell me whether a car is worth visiting. I created CarHunter to compare offers with my budget, likely repairs and the information still missing from the description. It tracks listing changes and reserves more expensive analysis for promising vehicles. My work covers the collection library, analysis graph, persistent records and contact scheduling. Collection and initial screening have run unattended on my infrastructure; the wider workflow is designed to prepare an informed purchase decision.

Collection across three marketplaces and initial screening ran unattended; deeper analysis and contact planning are implemented.

When my coding assistants work on the server, I want to follow their changes and intervene from my phone. CloudCLI already provides that web interface. I adapted the open source project for my workflow, particularly so a Claude approval request remains available when I return to an active session. My changes connect the displayed permission mode to the assistant’s actual mode, save decisions through its native project rules and keep favourites and task states consistent across browser connections.

Fork of siteboon/claudecodeui; rebuilt Claude approvals, native permission rules and improved session tracking.

Read the full story

Clothify starts with a garment photo and produces an image for presenting it on a person. Users submit their image in Discord, choose the garment type, build and view, then receive the result in the same channel. I developed that journey and connected the bot to PostgreSQL, n8n and an image model. Requests and processing states form a contract between components, letting generation evolve independently of the interface while distinguishing a finished image from one actually delivered to its user.

Working bot video; two garments shown from four saved angles in the demo.

View the demo Read the full story

I built this tool for dictating into an editor or conversation without opening another application or sending audio to an external service. It records speech, detects when a sentence ends and pastes the text into the active window. Whisper and Silero supply the models; my work connects them to everyday desktop use, handling pauses, unwanted transcription outputs and the transitions between listening, processing and sleep. Automatic pasting targets Ubuntu with X11, and recognition runs locally after the initial model download.

Local transcription, insertion into the active window and clipboard restoration.

Read the full story

For this NLP course project at ESIEA, I investigated how different networks identify a poem’s author among Baudelaire, Hugo and Rimbaud. I worked on text preparation, segmentation and the comparison of four architectures under a shared protocol. Segment predictions are combined into an attribution for the whole poem. The final report presents results on 32 test poems, errors by author and changes made during experimentation. The project also involved examining overfitting and the limits of the evaluation procedure.

87.5% on 32 test poems for both the MLP and LSTM.

In this symbolic AI project at ESIEA, a hunter must find gold and leave a dangerous grid without directly seeing neighbouring cells. I developed its reasoning from the teacher’s starting skeleton. The agent uses perceived clues to revise its picture of the world, evaluates possible actions and changes priorities once it holds the gold. The course’s simulator and interface make this loop observable. My work focuses on the beliefs, preferences and action-selection rules written in Prolog.

Belief revision, correction after a collision and a switch to returning once the gold is acquired.

How can an agent learn to drive from the images it receives? I connected the CarRacing simulator to an existing DreamerV3 implementation to explore that question. The algorithm learns a model of its environment and uses imagined trajectories to train its driving policy. I wrote the adaptation of observations, controls and episode signals, then ran roughly 41 hours of GPU training. The curves and video show progress within the tested setting: one circuit shared by training and evaluation.

About 41 training hours on an RTX 4060, with metrics and driving video from one circuit.

Read the full story

Reading a protected page can require several decisions: choose an engine, wait, preserve a session or try an interaction. I adapted the Scrapling framework to handle that orchestration through an MCP server used with my agents. The service starts with simple retrieval and can hand over a page that remains open when the agent needs to continue itself. My contribution covers this progression, navigation tools, profiles and diagnostics; the retrieval engines and parser come from the original project.

Scrapling adaptation with progressive retrieval, interactive pages and access regulation per domain.