Startwise

AI-powered startup ideation platform that turns a rough idea into a structured strategy through an interactive multi-agent Q/A workflow.

Overview

Startwise guides founders from idea to clarity by combining:

• A modern frontend experience (Next.js + TypeScript)
• A Django API backend for session/state orchestration
• Multi-agent reasoning with Google ADK + A2A
• OpenAI-powered iterative questioning and refinement with Firecrawl-backed research

The core flow is sequential and adaptive: each next question is generated based on the previous user answer.

Architecture

Tech Stack

• Frontend: Next.js, React, TypeScript, Tailwind, shadcn/ui
• Backend: Django, Python
• Agent Runtime: google-adk, a2a-sdk, uvicorn
• LLM Provider: OpenRouter via LiteLLM (meta-llama/llama-3.3-70b-instruct:free)

Repository Structure

Startwise/
├─ frontend/                 # Next.js application
├─ backend/                  # Django + ADK/A2A agent services
│  ├─ ideation/
│  │  ├─ agents/             # Question / Research / Formulator agents
│  │  └─ services/           # A2A client + orchestration logic
│  └─ manage.py
└─ README.md

Core Agent Flow

1. User submits business idea in frontend.
2. Django initializes ideation session.
3. QuestionAgent refines context (sync).
4. ResearchAgent produces strategic notes (streaming/SSE mode).
5. FormulatorAgent generates one next question.
6. User answers, and answer quality is evaluated.
7. Loop continues until completion criteria is met, then summary is generated.

Quick Start

1. Backend setup

From backend/:

uv pip install -r requirements.txt

Create env:

cp .env.example .env

Required env values:

• OPENROUTER_API_KEY=<your_openrouter_api_key>
• QUESTION_AGENT_MODEL=meta-llama/llama-3.3-70b-instruct:free
• RESEARCH_AGENT_MODEL=meta-llama/llama-3.3-70b-instruct:free
• FORMULATOR_AGENT_MODEL=meta-llama/llama-3.3-70b-instruct:free

2. Run backend services

From backend/, in separate terminals:

uv run python -m ideation.agents.question_agent
uv run python -m ideation.agents.research_agent
uv run python -m ideation.agents.formulator_agent
uv run python manage.py runserver 0.0.0.0:8001

If you want the product-audit retrieval store locally, start Qdrant from the repo root:

docker compose up -d qdrant

3. Frontend setup

From frontend/:

npm install
npm run dev

Set frontend env:

• NEXT_PUBLIC_BACKEND_URL=http://localhost:8001

API (Ideation)

Main endpoints:

• GET /health
• POST /init
• POST /respond
• GET /keywords/<session_id>/<question_index>
• POST /suggest
• GET /session/<session_id>
• DELETE /session/<session_id>
• GET /summary/<session_id>
• POST /summary-with-image/<session_id>
• POST /reset

Notes

• If agent services are unavailable, set USE_A2A_MOCK=true in backend/.env for local fallback.
• The A2A architecture is unchanged. The text agents now run through ADK's LiteLLM connector against OpenRouter instead of Gemini.
• For full backend details, see backend/README.md.