Zenquanta AI

Zenquanta AI is a premium multi-assistant workspace built with Next.js, TypeScript, Tailwind CSS, shadcn/ui, Supabase, and OpenRouter. Six assistant families share a workspace shell, model routing, memory, and a real plan ladder — closer to a desk than a chat box.

Most AI products collapse into a single chat input. That is a UX dead end. Zenquanta organizes assistants into families with their own surfaces, while sharing project context, memory, and a single billing layer underneath.

• Nova — general-purpose conversation
• Velora — long-form writing and structure
• Axiom — analytical and reasoning-heavy work
• Forge — coding and developer workflows
• Pulse — quick, low-latency responses
• Prism — image generation and visual flows

Tier-aware routing picks the cheapest acceptable model per request, given the user's plan, the assistant family, and the intent of the message. Provider adapters sit behind a single interface so the UI never knows which model is running.

Before a request hits a model, prompt precheck inspects intent and recommends an assistant if the user landed in the wrong family. It also flags requests that should be routed to image generation instead of text and vice versa.

Text and image flows are separate pipelines with their own routing, cost accounting, and surface. They share session context and memory but never silently swap with each other — Prism owns the image surface explicitly.

Conversation memory is scoped per assistant family but visible across the workspace. Usage tracking is split into two layers: raw / admin cost (what providers actually charged) and displayed user-facing usage (what the plan shows). Keeping these separate is what makes a real billing system possible.

Plans are explicit: tier defines model access, monthly usage, image quotas, and which assistants are available. Manual plan requests route through admin activation, so paid plans never quietly auto-grant.

• Admin dashboard for plan activation and abuse handling
• User dashboards with per-assistant usage breakdown
• Public assistant pages for discovery
• Working notes shown during streaming so the user sees thought, not just output

• Keeping raw provider cost and displayed user usage from drifting apart
• Streaming-aware fallback when a provider goes down mid-response
• Handling memory across multiple assistant families without leaking context
• Designing a plan ladder that is honest about what is included without burying the user in tier matrices

• Long-running agents that keep state across sessions
• Background tasks queued from any assistant
• Team workspaces with shared memory and shared usage
• Self-serve plan changes with proration
• Open assistant SDK for community-built families