Cap chart size: max-width 880px and max-height 460px so charts have natural proportions

Wider left column (480-580px) + capped schema height with scroll

Fluid layout (no max-width) to remove dark side bars

UX premium: chart badge, SQL syntax highlighting + copy, smart placeholder, drag highlight

Cap to top 30 categories, hide gr.File X button, single-row buttons, no duplicate title

Polish: 4-step pipeline progress, large empty state w/ icon, subtle entrance animations

Bigger rectangular layout: 1600x920 max with 16:9 feel

Dynamic suggested questions based on actual dataset columns (no more Titanic-only)

Layout: 2-column split (left=data+question, right=results)

Result narration (Qwen-powered) + UI polish for upload area + dark-mode fixes

SVG: standalone download (XML prolog, white bg, fixed dims) + Plotly first

Dark-mode fixes: visible title/chip, hide gr.File uploaded box, transparent question group

Remove hf_oauth + LoginButton — they conflict with native ZeroGPU auth

UX: schema below dataset / above textbox, single-shot Ask, Clear=dataset only

Clear button now resets the upload file too

Add LoginButton + hf_oauth: required for ZeroGPU Pro quota in iframe context

UX: CSV/SVG download links, schema preview on load, cleaner SVG theme

ZeroGPU best practice: load models at module level (cuda), inference only inside @spaces.GPU

Remove hf_oauth/LoginButton, pre-download models in build phase

Add gr.LoginButton for HF OAuth

Sequential model load/unload: keep peak VRAM at ~14GB

Add quota banner with HF login link in header

Fix: GPU duration 120s (max), reactive empty/ready state from agent.list_tables()

UI v2: progress streaming, demo dataset, suggestions, polished CSS for Gradio 5

Disable API schema introspection (gradio_client bug w/ Dict[str,Any])

Initial deploy: Apple/Claude design, DuckDB, 3 trained LoRAs