Anyesh/wardrowbe

wardrowbe – Self‑hosted AI‑powered wardrobe manager

What it is – A full‑stack web application that lets you photograph your clothes, stores them in a searchable catalog, and uses a multimodal LLM (via OpenAI, Ollama, LocalAI, etc.) to tag each item and generate daily outfit suggestions based on weather, occasion and personal preferences. All data lives on your own hardware; the AI component can be run locally (Ollama) or through any OpenAI‑compatible service.

Core features (as listed in the README)

Photo‑based catalog – Upload pictures; the AI extracts colour, pattern, style, and optionally removes the background.
Smart outfit recommendations – Multimodal LLM suggests what to wear, taking weather (Open‑Meteo) and user‑defined occasions into account.
Scheduled notifications – Daily suggestions can be pushed via ntfy, Mattermost, or email.
Family accounts – Separate wardrobes for multiple household members.
Wear tracking & feedback – Log when you wear an item, rate the outfit, and give comments.
Analytics dashboard – Visualise colour distribution, most‑worn pieces, items that never get used, etc.
Fully self‑hosted – Docker‑compose (or Kubernetes) deployment; no SaaS lock‑in.
AI‑agnostic – Works with any OpenAI‑compatible API: Ollama (free, local), OpenAI, LocalAI, or other hosted models.
Internationalised UI – UI translated into eight languages via next‑intl.

How it works (architecture from the README)

Frontend (Next.js 14 + React Query)  <--->  Backend (FastAPI + async SQLAlchemy)
        │                                 │
        │   PostgreSQL (wardrobe data)    │   Redis (job queue)
        │                                 │
        └───► Background worker (arq) ──► AI service (Ollama/OpenAI/etc.)

Upload – The frontend sends the image to the FastAPI backend.
Background job – An arq worker picks up the request and calls the configured vision model (e.g., gemma3 via Ollama) to extract tags and optionally remove the background.
Storage – Tags and metadata are saved in PostgreSQL; the image (or its trimmed version) is stored on the host volume.
Recommendation engine – A scheduled job (or on‑demand request) calls the text model (same or different) with the current weather (Open‑Meteo) and user preferences, returning a natural‑language outfit suggestion.
Delivery – Suggestions are sent to the user via the chosen notification channel.

Tech stack (explicitly mentioned)

Layer	Technology
Frontend	Next.js 14, TypeScript, Tailwind CSS, shadcn/ui, TanStack Query
Backend	FastAPI, async SQLAlchemy, Pydantic, Python 3.11
Database	PostgreSQL 15
Cache / Queue	Redis 7 (used by `arq` workers)
Background jobs	`arq` (async job queue)
Auth	NextAuth.js (dev credentials or OIDC)
AI integration	Any OpenAI‑compatible API (Ollama, OpenAI, LocalAI, etc.)
Containerisation	Docker Compose (multi‑arch images) – optional Kubernetes manifests

Getting started (quick‑start summary)

Install Docker + Compose and, if you want a local LLM, install Ollama and pull gemma3 (or another multimodal model).
Clone the repo and copy .env.example → .env.
Edit .env to point at your AI service (Ollama URL or OpenAI key) and set secrets (SECRET_KEY, NEXTAUTH_SECRET).

Run the production compose file:

docker compose pull
docker compose up -d
docker compose exec backend alembic upgrade head   # migrations

Open http://localhost:3000 for the UI and http://localhost:8000/docs for the API docs.

For development, add -f docker-compose.dev.yml to get hot‑reload and run the frontend locally with npm run dev.

Typical use cases

Personal wardrobe digitisation – Scan every garment once and let the AI auto‑tag it.
Daily outfit planning – Receive a weather‑aware suggestion each morning.
Family sharing – Keep separate closets for spouses, kids, etc., all under one self‑hosted instance.
Data‑driven decluttering – Use the analytics view to see which colours or items you never wear.

Limitations & considerations

AI quality depends on the model – Free local models (e.g., gemma3) may be less accurate than OpenAI’s vision models.
Hardware requirement – At least 4 GB RAM; larger multimodal models need more memory.
Background‑removal optional – Requires rembg or an external HTTP service; otherwise the button returns a 501.
Privacy of location fallback – Enabling NEXT_PUBLIC_ENABLE_IP_LOCATION_FALLBACK sends the user’s IP to a third‑party service.
No mobile app yet – The README mentions an upcoming Play Store entry; currently the UI is web‑only.

Who might find it useful

Tech‑savvy individuals who want a privacy‑first digital closet.
Small families looking for a shared wardrobe manager without relying on cloud services.
Developers interested in building AI‑augmented personal‑assistant tools and needing a ready‑made example of vision‑plus‑text integration.

Bottom line – wardrowbe is a genuine, self‑hosted application that combines a modern web stack with multimodal LLMs to automate wardrobe organization and outfit recommendation. It is fully documented, containerised, and configurable to run with any OpenAI‑compatible backend, making it a solid example of an AI‑enhanced personal‑productivity tool.