The idea

Personal data usually lives in separate apps that don't talk to each other: a bank app, a notes app, a task list, each behind its own account. Everyday puts all of it in a single self-hosted account, encrypted at rest, that I run and control myself.

What it does

The app covers a few areas of daily life. Each one works on its own, and the assistant can act across all of them.

• Finances: net worth tracking, budgets, savings goals, portfolio rebalancing, and expense analysis from imported bank statements.
• Meals: a recipe book written in Markdown, with AI-generated cover images and a log of what I've cooked.
• Wiki: a personal knowledge base with links between notes.
• Tasks: to-dos wired to a Gantt and Kanban planner, alongside a dated journal.
• Wardrobe: an inventory of clothes with an AI styling assistant and image generation.
• Assistant: a chat companion that reaches across everything above.

Why it's agentic

The assistant can act on the app directly through a set of tools, built on the Model Context Protocol, so it reads and edits the actual data: finances, notes, tasks and the rest. It remembers earlier conversations, sends a reminder when something is worth flagging, and works from the web app or from Telegram.

Under the hood

The whole codebase is TypeScript, front and back. Each part runs in its own container, and releases go out through a CI pipeline.

• React 19 and Vite on the front end, Fastify and MongoDB on the back.
• Personal fields are encrypted one by one at rest, using AES-256-GCM.
• Redis runs the background jobs, and files are stored on Cloudflare R2.
• AI is bring-your-own-key: each user configures their own provider, and the keys are kept encrypted.
• Everything ships in Docker, with releases and deploys running through GitLab CI/CD.

Status

It's a personal project under active use and maintenance, self-hosted and already running in production for my own use. I add features gradually, fix bugs, and keep the software up to date.

The idea

On most platforms, making a character means filling in a long form: physical description, backstory, opening line, tags, and an avatar to find or draw. It's tedious, and plenty of people give up before the character is any good. Amadeus takes a name and a few personality traits, writes the rest, and generates a matching portrait. Everything it produces is a first draft, and it all stays editable afterwards.

What it does

The app has two parts: creating characters, and chatting with them.

• Guided creation: from a name and a few traits, it drafts a description, a backstory, greeting lines and tags.
• Portrait generation, cropped automatically around the face into avatar, thumbnail and full-size versions.
• An image studio to retouch a portrait from a written instruction, preview the change, then keep or discard it.
• Roleplay chat that keeps its full history, with a system prompt assembled from the character's own profile.
• Search across characters and tags, plus suggestions of characters that are semantically close.
• Public and private characters, likes, favourites, moderation reports, and export to a PNG character card.

The agentic part

Creating a character runs as a series of model calls. One handles the appearance, another the background, then the greeting, then the tags. Each step has its own instructions and cleanup, and each one feeds the next. The text that comes out is what gets turned into the prompt for the image model. The pipeline also calls a number of services by itself: image generation, face detection, embedding computation, vector search, object storage and search indexing. The model providers are chained, so if one goes down the next takes over without dropping the conversation. A word on "agentic", since it's an easy term to oversell: this is orchestration with fallbacks and background workers. The model doesn't set its own goals; the order of the steps and the tool calls are all written in code.

Technical approach

React and TypeScript on the front, FastAPI on the back. Each concern runs in its own container, so a slow or failing piece doesn't take the rest down with it.

• Front end: React, TypeScript, Vite, Tailwind, React Query and Zustand.
• API: FastAPI, fully async, split into endpoints, services and repositories.
• Storage: PostgreSQL with pgvector, MongoDB for conversations, Redis for cache and the task queue.
• Background work: Celery workers for embeddings, counters and index rebuilds.
• Search: Meilisearch for text, and a separate FAISS service over gRPC for vectors.
• Infrastructure: a Docker Compose setup per environment, Loki and Promtail for logs, and S3 with signed CDN URLs for images.

Current status

Development has stopped. I don't align with the values of this market, and I don't believe a conversational AI can be made reliably safe for everyone, a limitation intrinsic to the technology.