SubsGPT uses the SBS_AI backend for authentication. Log in with the email and password your administrator gave you. If you don't have an account, ask the administrator to send an invite — the link sets your password directly.

Open Proposals → New. The proposal builder is a six-step wizard: Setup → Spec → Design → Map → Generate → Summary. Your work auto-saves as you go (locally first, then to MongoDB on a 2-second debounce — visible as the "Saving…" / "Saved" chip in the header).

On Step 0 (Setup), start typing a city, address, or region in the location field. The ArcGIS geocoder returns suggestions after 3 characters. Selecting a result auto-fills latitude, longitude, country, region, and city — and triggers auto-suggestions for frequency (50 / 60 Hz), grid code, environmental class, and BIL on the next step.

Step 1 (Spec) covers protection scheme, fault level, SCADA protocols, BIL, environmental class, and arc flash. The arc flash category is calculated automatically from fault level + protection scheme using simplified NFPA 70E energy bands; click the override toggle if you want to set it manually. Auto-set fields show a small chip so you know what the system filled in.

Step 3 (Map) embeds the 3D ArcGIS scene. Click Draw Site and click points on the map to outline a substation footprint — the polygon's bounding box becomes the site dimensions. From the toolbar you can generate a substation layout for each site (using the neural-net layout model — pick which checkpoint to use in Settings → Neural Net), place wind turbines and transmission towers, and toggle ESRI feature layers (transmission lines, substations, power plants).

Step 4 (Generate) shows a card grid of the 19 document and drawing types SubsGPT can produce — Proposal, Technical Spec, Single Line Diagram, Site Layout, etc. Each one is mapped to an agentic config in the backend. Click Generate on a card to run the pipeline; the resulting document opens in the preview modal where you can read and export it.

Click the Chatbutton in the proposal-builder header at any step. Design Chat is RAG-grounded engineering Q&A backed by the rural-substation-advisor agentic pipeline (3 agents: searcher → analyst → engineer-formatter). Pick a standards corpus from the dropdown, ask a question, and get a structured markdown answer with inline citations. Each turn takes ~30–60 s on Ollama and is independent (it doesn't see prior turns) — but the conversation is preserved in the proposal blob and survives reload.

Click Verify Standards next to Chat. The drawer lists every site in the proposal that has a generated layout. Pick a site and a standards corpus, then click Run verification. The verify-layout agentic pipeline runs (clearance-extractor → geometry-analyser → compliance-reporter) and returns a structured compliance report — a sortable PASS / REVIEW / FAIL table with cited source documents and pages, and an overall verdict.

Open Settings. The tabs cover the LLM backend (Ollama / OpenAI / Anthropic / Azure / Bedrock / none), retrieval weights, the neural-net checkpoint preference (which.pt the Map Designer uses for layout generation), document-to-config mappings, and import / export. The neural-net preference is browser-local — pick a default per workstation.