User Guide
Everything you can do on GPU Clouds, in one place — what each tool is for, who can use it, and how the pieces connect.
Accounts & access
Some tools are open to anyone; others need an account.
No account needed
GPU Sizing Advisor — chat, get a recommendation, request a Build Plan by email.
Account required
War Room, My Brain, your project dashboard, and anything tied to your history.
Register at /auth/register or sign in at /auth/login. If a teammate invites you into a War Room session, follow their invite link first — it carries you straight into registration and drops you into the exact lens they assigned you.
GPU Sizing Advisor /gpu-sizing/
A conversation with Atlas, a hardware-sizing advisor, that ends in an itemized cost breakdown for the compute your workload actually needs.
What it asks about
Atlas gathers what it needs a couple of questions at a time, not a form: workload type (training, inference, RAG, vision, batch…), model or framework, data scale, concurrent users, latency tolerance, security or compliance constraints, cloud vs. on-prem preference, budget sensitivity, and your timeline and in-house team capability.
What you get back
A primary configuration and a budget-friendly alternative, each with a line-item cost breakdown, sourced against current reference pricing across GPUs, CPUs, NPUs, and cloud instances. Every price carries a stated ±20% variance and a freshness date — treat it as a planning estimate, not a quote.
Free AI Build Plan. Once you have a recommendation, you can request a one-page build plan by email — no account needed, just an email address. It's delivered to your inbox along with a qualification summary, and (if you want) an invitation to book a discovery call.
If you're signed in when you use the Sizing Advisor, anything durable it learns about you and any recommendations it generates also flow into My Brain — anonymous visitors get the sizing tool with none of that follow-through, by design.
War Room /war-room/
Requirements capture built for more than one stakeholder — everyone states their own priorities, and the platform surfaces where those priorities actually conflict.
Start a session
Give it a title and, if you have one, a North Star — the plain-English goal everything else gets measured against. From there, fill in whichever of the four lenses is yours.
Business
Outcome, users, competitive signal.
Tech
Hard constraints, integrations, deployment target.
HR
Affected roles, supervision, escalation.
Finance
Budget envelope, ROI horizon, cost today.
Each lens is a mix of priority sliders and a few direct questions. Don't own a lens yourself? Invite the person who does — they get an email link that takes them straight to registering (or logging in) and lands them on exactly that lens, nothing else.
The Friction Agent
Runs automatically every time a lens is saved, and only flags real contradictions — not just differing priorities. A Finance budget cap that makes Tech's latency target physically unreachable, for instance, comes back with a concrete suggested trade-off. Resolve it (leave a note on how you settled it) or dismiss it if it doesn't apply.
Compiling the Blueprint
Once at least one lens is filled, compile. The Architect Agent turns everything captured — including any negotiated friction — into a Blueprint:
- An executive summary and a Fragility Index (0–100: rock-solid managed stack → experimental)
- Three Swarm Options — Speed Demon, Sovereign Stable, Efficiency Lean — each a different cost/latency/model trade-off
- A Business Capability Map, a Tech Stack & latency budget, an HR human-in-the-loop workflow, and a Finance total-cost-of-intelligence view
- New Recommendations specific to this Blueprint — see My Brain
Compiling also creates a Project, so the Blueprint stays attached to something you can keep coming back to.
My Brain & Recommendations /brain/
A private memory that carries what the platform learns about you between conversations, plus a running list of architecture recommendations generated from your own Blueprints and build plans.
How memory works
As you use War Room and the Sizing Advisor, durable facts worth remembering — a preference, an infrastructure detail, a compliance requirement — get proposed as pending memories. Nothing is used automatically: you accept or reject each one on the My Brain page. Only accepted memories ever get fed back into a future conversation, and you can see exactly when each one was last used.
Delete any single memory, or erase everything in one action. Export the whole thing as a JSON file whenever you want a copy for yourself.
Recommendations
Every time a War Room Blueprint compiles, or you generate a Sizing Advisor build plan while signed in, the Recommendation Engine checks it against a curated library of best practices and known pitfalls, and surfaces what actually applies — grouped by how urgently it matters:
Each finding comes with why it matters, the evidence it's grounded in, and a suggested next step. Accept it, defer it for later, or dismiss it — dismissed findings won't keep resurfacing unless something material about your project changes.
Privacy controls
Two settings live on the same page: memory mode (Private & Persistent, or Off — stop new memories from being written at any time) and an opt-out for the collective signal. That signal, when it's on, only ever records whether a recommendation type was accepted or rejected in de-identified, banded form — never your content, never anything traceable back to you or your project.
From requirements to a delivered system
For anyone who wants GPU Clouds to actually build and run the thing, not just plan it — the guided engagement that follows a Blueprint or a direct conversation.
Discovery
A conversation with Aria, who works through your problem, your end users, expected scale, what systems already exist, constraints, integrations, and timeline — then hands you a structured summary to confirm before moving on.
Scoping
You choose an engagement model (Full Handhold, Build & Deploy, Build Only, Advisory, or Co-Build), a hosting preference (Managed, Client Cloud, On-Premise, Hybrid, or Edge), and a support tier. Aria lays out exactly what you'll need to provide before work starts — credentials, data, approvals, whatever applies.
Proposal
A phase-by-phase plan with pricing, a cost breakdown, called-out risks, and a payment schedule tied to each phase.
Execution
Once approved, work is tracked phase by phase with progress updates, and each phase is verified before handover to the next.
For administrators /admin/
Platform oversight — visible only to accounts with the admin role.
- Projects & Agents — status and history across every engagement
- Architecture Pattern Library
/admin/patterns— the curated best-practices and known-pitfalls the Recommendation Engine draws from; add, retire, or update entries - Collective Signal
/admin/signals— de-identified accept/reject outcomes, only ever shown once enough accounts have contributed to a given category. Below that threshold, it shows nothing — that's intentional, not a bug.
Notes & limits
- War Room updates (friction points, invite status) refresh on a short interval rather than pushing instantly — give it a few seconds, or refresh the page.
- The Sizing Advisor works without an account by design, as a public lead-in tool — sign in first if you want its output to reach My Brain.
- Every cost figure anywhere on the platform is a planning estimate with a stated variance and a pricing date attached — confirm current pricing before committing a budget.
- My Brain and Recommendations are being rolled out gradually — if you don't see My Brain in your navigation yet, it isn't enabled for your account.