LIVE PRODUCT

Nexus IA

Nexus IA — Centralized artificial intelligence

Use now Architecture

nexus.expostacker.com.br

live preview

Tech stack: FastAPI Python React Vite AI Groq

5
Nexus models: 3
AI providers: 4.7s
Average response time

Client

ExpoStacker

Sector

Artificial intelligence and productivity

Timeline

3 semanas · 2026-08

Live product

Use now

Results

Client-reported results during rollout

25/25 Chat tests OK

Baseline: 11/25 before event loop fix

Fonte: NEXUS audit

2.1s Landing load time

Baseline: without asset cache (estimated)

Fonte: Local test

15+ Media queries for responsiveness

Baseline: adaptable layout from 320px to 1200px

Fonte: CSS audit

Client-reported data

The challenge

Nexus IA is an artificial intelligence SaaS that centralizes multiple models in a commercial dashboard. The platform offers five model profiles, context-aware chat, code generation and execution, workspace with terminal and an agent capable of creating files and editing projects.

Equipes e profissionais desperdiçam tempo alternando entre múltiplas ferramentas de IA, corrigindo respostas com acentuação quebrada e lidando com limites de uso de APIs individuais, o que reduz produtividade e aumenta custos operacionais.

The approach

We started from a simple dashboard and evolved into an architecture with latency-based routing, key pools per provider, persistent cache and automatic fallback. The front-end keeps the ExpoStacker visual identity with mobile responsiveness and accessibility.

The solution

Dashboard with five Nexus models and automatic task-based routing
Code assistant with generation, explanation and review
Workspace with agent chat, terminal and quick actions
URL research with context for page-based answers
User plan, limit and cost control
Commercial landing with SEO, schema.org, sitemap and robots.txt

Timeline

Como o projeto foi conduzido, do mapeamento ao deploy. How the project was conducted, from mapping to deploy.

Semana 1

Descoberta

AI provider and key mapping
Definition of Nexus models and personas
Routing and cache architecture

Semana 2

Construção

FastAPI with chat, code, crawl and agent
Commercial dashboard with landing and plans
Vite workspace with login and authentication

Semana 3

Validação

Tests of all models and prompts
SEO, sitemap and responsiveness audit
Encoding and provider fallback adjustments

Methodology

Como o projeto foi conduzido, do mapeamento ao deploy. How the project was conducted, from mapping to deploy.

Continuous refinement sessions

Load and latency tests per model

SEO and accessibility audit

ExpoStacker component standards

Architecture

Usuário

Navegador

FastAPI

Nexus IA

Provedores

Groq / Cerebras / Gemini

Workspace Vite

Painel Web

Cache Redis

Banco SQLite

→ FastAPI orchestrates chat, code, crawl and agent
→ Router picks provider based on latency and quota
→ Round-robin key pool per provider
→ Persistent prompt and response cache
→ React Vite consumes the API via /api prefix

Conceptual diagram — does not reflect real implementation

Tech stack

Technical choices and what was left out.

Backend

FastAPI

Why: Async Python for external APIs

Rejected alternative: Django (heavier)

SQLite

Why: Local database without external infrastructure

Rejected alternative: Postgres (future scale)

httpx

Why: Async client for external calls

Rejected alternative: requests (synchronous, blocking)

Frontend

React + Vite

Why: Fast SPA for the workspace

Rejected alternative: Next.js (overkill for MVP)

Tailwind

Why: Tokens consistent with ExpoStacker

Rejected alternative: Plain CSS (slower to prototype)

AI

Groq

Why: Low latency and competitive cost

Rejected alternative: OpenAI (higher cost)

Cerebras

Why: Large models without local infrastructure

Rejected alternative: Local Ollama (excluded by scope)

Gemini

Why: Fallback with excellent Portuguese

Rejected alternative: Claude (no key in project)

Decision log

What I chose, what I rejected, and why.

Use provider routing with fallback

Why: Prevents failures due to unavailability or quota of a single external provider.

Rejected alternatives: Single provider simplifies code but creates a single point of failure.

MVP with SQLite instead of Postgres

Why: SQLite removes external database dependency and speeds up initial deploy.

Rejected alternatives: Postgres offers more scale, but requires extra infrastructure setup.

Dashboard in vanilla HTML/CSS + React Vite for workspace

Why: Fast and light landing; complex interactive workspace as React SPA.

Rejected alternatives: Everything in React would increase landing bundle and hurt SEO.

FASSI.AI

Intelligent routing pipeline

User sends the message

Classifier identifies the task

Router picks the best Nexus model

System tries providers by score order

Response is cached and recorded for cost

Engagement mechanics

Progressive plans

Free, Starter, Pro and Business unlock models and limits based on usage.

Active workspace

Projects and files keep the user returning to the environment.

Models per task

Nexus Auto, Fast, Standard, Advanced and Code teach the user to pick the right tool.

Security and access

Security and compliance

External API keys never exposed in the front-end
JWT authentication with secure cookies
Rate limit per route and per plan
CORS restricted to allowed domains
Dev environment flagged for future hardening

Lessons learned

Not everything went as planned.

— Sharing global async HTTP clients between synchronous requests causes event loop errors. Creating httpx clients per request solved timeouts and external provider encoding issues.