jellai-showcase/README.md

JELLAI

An AI-native TV experience layered over an existing Emby media server. Not a library you browse — a channel that's already on, plus a concierge you talk to. A single tool-calling model runs the whole loop: it drives the gateway tools (search, schedule, playback handoff) and talks to you from a personality prompt. Always-on chat, TV-style time-block scheduling, mood awareness, and preference learning seeded from your real watch history.

Status: v1 — working prototype. TV-first, single user. Built on the Emby you already have (one server, one source of truth — your watch history is the personalization seed).

What it feels like

You don't open JELLAI. It's already there — your media server, just with the lights turned down and a concierge in the corner. A TONIGHT strip runs along the bottom, and a 💬 door waits bottom-right. Nothing covers what you're watching.

Tap the door and the concierge panel slides in from the right — video keeps playing. Tell it how you feel, not what to search for: "build me a chill TV night — something funny and easygoing." It talks back in its own voice, then builds a real lineup into the strip: time blocks, in order, drawn from your actual library. Don't like it? [yes] [swap] [keep] — one word steers the whole night. Click any block to start playing from there, in Emby's own player.

The strip is the spine. Ask once, and the rest of the night schedules itself — then re-schedules live every time you change your mind. That's the whole product: a channel that's already on, tuned to you, that you can talk to.

These are real screenshots — a live themed Emby client driving the running gateway and a local language model. No cloud LLM in the loop.

How it works

Emby web UI  ──CSS/JS inject──▶  concierge panel ──WebSocket──▶  gateway ──▶  local LLM (Ollama)
(your server)                     (overlay theme)                (tool loop)    tool-calling
       ▲                                                            │
       └──────────────── Emby REST API (/Items, schedule, playback)─┘

• Frontend (v1): a concierge panel + schedule strip injected into the Emby web client via CSS/JS. It couples to almost nothing in Emby's DOM — it only appends its own fixed-position nodes — so it survives Emby updates. Playback always delegates to Emby's own player. The eventual target is a custom SPA over the Emby REST API.
• Gateway: a small Python WebSocket server mediating between the UI and a local model. It runs a tool-calling loop against one stock instruct model — no fine-tune unless a measured gap demands it.
• Tools: search, TV-style scheduling, and watch-state queries against the Emby API. The model sees JSON tool schemas and composes them (e.g. "movie night" → a movies-only lineup).

Why build on Emby instead of a new frontend

The original plan was a dedicated themed media app. We reversed it and built on the existing Emby:

• Single source of truth — no duplicate library scan, no second metadata DB.
• Watch history is the personalization seed — per-user learning is the whole premise, and that history already lives in Emby. A fresh install starts cold.
• Emby is enough — a full documented REST API, plus community CSS/JS injection for the overlay.

Status & scope

Working today: the concierge loop, live TV-style scheduling into the strip, mood-aware search with type/year/rating/sort filters, and one-word lineup steering. Deliberately out of v1: content acquisition, multi-user profiles, and model fine-tuning — added only against a measured need.