Perspektiv

May 13, 2025

Concept & Design

Perspektiv is an AI-powered debate tool that transforms a single prompt into a structured, multi-perspective report using a format like Pro/Con/Synthesis. The goal was to build and ship a functional SaaS product—fast, cheap, clean. Nothing extra. It needed to:

Let users input a topic and select a format
Use GPT-4o to generate structured arguments
Stream results live and support PDF export
Keep infra + dev time minimal

What I Built

A clean monorepo (Next.js frontend + FastAPI backend) tied together with Celery for async jobs, PostgreSQL for structured data, and Clerk + Stripe for auth and billing.

Key Features

Real-time SSE streaming
Debate templates (Pro/Con/Synth)
PDF export from UI
Stripe subscription management
Clean dev experience via Cursor IDE and Docker

1. Quick Tech Spec

| Layer | Choice & Why | | ------------- | ------------------------------------------------------ | | Frontend | Next.js + Tailwind – slick DX, SSR, easy Vercel deploy | | State | React Context + useReducer – minimal boilerplate | | Backend | FastAPI (Python) – async, type‑hinted, automatic docs | | LLM | GPT-4o via openai SDK – abstracted provider file | | Queue | Celery + Redis – triggered only if job > 5s | | DB | Postgres – users, debates, messages | | Storage | S3 – for transcripts & PDF exports | | Auth | Clerk.dev – JWT tokens | | Billing | Stripe – tiered subscriptions | | Hosting | Vercel (FE), AWS Fargate (API + Worker) | | Observability | Sentry + CloudWatch (basic setup) | | Targets | ≤15s p95 latency · ≤ $0.05 cost per debate |

2. Minimal Architecture

Browser → Next.js → FastAPI → GPT‑4o        (short debates)
                   │
                   └─ Celery/Redis → GPT‑4o  (long debates)
                   │
                   ├─ Postgres
                   └─ S3 (exports)

SSE streams responses directly to frontend.

3. Core User Journey

Sign-up via Clerk OAuth
Dashboard shows usage + past debates
New Debate → select template, submit prompt
Loading bar shows real-time streaming
Result view with Pro/Con/Synth tabs
Export as PDF
History page → revisit old debates
Account → Stripe portal for plan changes

4. Build Guide (Cursor on macOS)

Phase 0 · Prereqs

Bash

brew install redis awscli
pyenv install 3.12.3 && pyenv global 3.12.3
npm install -g pnpm

Clone the GitHub repo: perspektiv-mvp

Phase 1 · Monorepo Layout

/ (root)
  frontend/          # Next.js
  backend/           # FastAPI + Celery
  docker/            # Docker & Compose files
  .env.example
  package.json
  turbo.json

Phase 2 · Backend Setup (FastAPI)

Bash

cd backend
python -m venv .venv && source .venv/bin/activate
pip install fastapi[all] uvicorn psycopg[binary] redis celery "openai>=1.12" python-dotenv boto3

Models: User, Debate, Message (SQLAlchemy)
/debates routes: POST to create, GET/stream to stream results
LLM service wraps GPT-4o API
Celery task runs long debates and stores results
Auth via Clerk JWT introspection
Stripe webhook syncs plan + quota info

Phase 3 · Frontend Setup (Next.js 14)

Bash

pnpm create next-app frontend --ts --tailwind

Auth with Clerk → wrap <ClerkProvider>
Routes:
- /dashboard: list debates, start new
- /new: input topic, pick template
- /debate/[id]: stream + view result
Use EventSource to receive real-time LLM output
Export via React‑PDF or backend-generated PDFKit

Phase 4 · Local Dev Orchestration

Yaml

docker-compose.yaml
services:
  db:
    image: postgres:16-alpine
    ports: [5432:5432]
  redis:
    image: redis:7-alpine
    ports: [6379:6379]

Set DATABASE_URL and REDIS_URL env vars.

Phase 5 · Deploy

Frontend: vercel --prod
Backend/Worker: AWS Fargate
- FastAPI task (port 8000)
- Celery task (no port)
- IAM role for S3 + SecretsManager
Infra:
- RDS Postgres (free tier)
- Redis (AWS ElastiCache or Upstash)
- S3 (presigned export links only)
Webhooks:
- Stripe and Clerk point to api.perspektiv.ai/webhook/*

Phase 6 · Polish & Launch

Seed DB with Pro/Con/Synth templates
Rate-limit free plan (10 debates/month)
Sentry in both FE + BE
robots.txt, landing copy
README for devs

Final Thoughts

I built and deployed this in a 7-day solo sprint:

2 days backend
2 days frontend
1 day infra setup
2 days polish + QA

It’s fast, clean, modular, and designed to scale just enough to get feedback before iterating. If you want to fork it, use it, or build on it—feel free to reach out.

Thanks for reading. You can explore the project at perspektiv.ai.