AI projects
Practical projects built around AI.
Recovering context, working with a document or accessing a clearly defined action: I am interested in AI when it becomes part of a useful tool. These applications and integrations show how that works, with a clear scope for each project.
Published on iPhone and iPad
Hephaistos
Personal AI, from mobile app to server.
Hephaistos brings Jarvis conversations, personal memory, sourced research and documents into one app. Users can return to their work without scattering context across separate interfaces. Voice and automations extend those uses.
My contribution : An application designed and published on the App Store under my name.
Practical uses
- Recover a preference, an idea or a working context.
- Ask a question about a document and inspect useful references.
- Research the web with sources that can be checked.
- Schedule a routine and follow its execution.
Architecture
Application and services
A Flutter interface communicates with a FastAPI backend and Supabase services. Conversations use HTTPS and WebSocket connections.
Choice of compute
Users choose Hephaistos Cloud or a compatible personal server. The setup guide offers macOS Apple Silicon, Linux and Windows: MLX on Mac, or Vulkan when available and CPU paths on Linux and Windows.
Scope and limits
- Jarvis requires a server path, including in personal mode. The selected mode determines where inference runs.
- Personal nodes also use account, pairing and remote-access coordination services.
- No public Android release is claimed here. Direct-message encryption does not describe the encryption of Jarvis conversations.
Integrated into this portfolio
Ask Logan
An assistant for exploring my background.
Ask Logan answers questions about my experience, skills and the projects shown here. It responds in English or French using portfolio information, helping visitors find a relevant detail.
My contribution : Design and integration of this website’s assistant.
Practical uses
- Identify the environments I have worked in.
- Find a particular skill or understand a project.
- Explore my background without reading every page.
Architecture
Documented context
The dialogue calls the portfolio API. Gemini requests run on the server with context assembled from public website data.
Response continuity
Local answers provide a fallback when the provider is unavailable. The interface distinguishes them from Gemini responses. Requests have a maximum duration and volume controls.
Scope and limits
- The assistant answers about this website. It does not administer systems or access company records.
- Answers cover the documented portfolio information and the working principles I describe here.
- Input and output checks constrain responses without guaranteeing accuracy. The source pages remain the reference for verifying a fact.
Public source code · desktop plugin
Secret Actions for Obsidian
Connecting MCP clients to approved actions.
Secret Actions allows REST or MCP clients to trigger predefined HTTP calls while credentials remain in Obsidian’s native secret storage. Clients use named actions and declared parameters, without general access to credentials.
My contribution : Development of a plugin published under my GitHub account.
Practical uses
- List GitHub repositories through a configured action.
- List Vercel or Supabase projects using the public samples.
- Define an action’s destinations, parameters and confirmation policy before enabling it.
Architecture
MCP interface
Three tools list actions, check their readiness and execute a selected action. The TypeScript plugin integrates with Obsidian’s Local REST API plugin.
Bounded execution
Schemas validate parameters. The Node transport applies destination and redirect restrictions; native storage supplies credentials, and results are filtered before returning to the client.
Scope and limits
- An independent desktop project requiring Local REST API. Users must configure and enable each action.
- The source is public under the MIT licence. Its documentation describes the REST/MCP contract and execution limits.
- An approved action may change an external service. Permissions and confirmation policies must match its effect.
Related software project
Le Barger, a restaurant ordering application.
Le Barger takes orders for a burger restaurant and connects the customer journey with restaurant operations: menu, ordering, tracking and fulfilment. It demonstrates the integration of a customer application, data and restaurant workflows.
Explore Le BargerPrinciples behind these projects.
- Start with an identifiable need and an understandable use.
- Define what the tool may read or execute.
- Provide a way forward when a service is unavailable.
- Document limits and keep sources available.
An idea to explore together?
I can walk you through these projects, their technical choices and the questions they help address.
Discuss a project