Build plan: Market Research Agent

Context

Build a local demo of an AI agent that helps solo founders discover their next product idea. The founder describes their situation in natural language; the agent conducts one-shot market research via web search and delivers an interactive dashboard report.

Requirements Summary

Architecture

┌─────────────┐     POST /research     ┌─────────────┐
│  React SPA  │ ─────────────────────► │  FastAPI    │
│  (Vite)     │                        │  Backend    │
│             │ ◄── JSON Report ────── │             │
└─────────────┘                        └──────┬──────┘
                                              │
                       ┌──────────────────────┼──────────────────────┐
                       ▼                      ▼                      ▼
                ┌─────────────┐       ┌─────────────┐       ┌─────────────┐
                │ DuckDuckGo  │       │ DuckDuckGo  │       │ DuckDuckGo  │
                │ Trends      │       │ Competitors │       │ Pain Points │
                └─────────────┘       └─────────────┘       └─────────────┘
                       │                      │                      │
                       └──────────────────────┼──────────────────────┘
                                              ▼
                                       ┌─────────────┐
                                       │  Kimi K3    │
                                       │  Synthesis  │
                                       └─────────────┘

← scroll the diagram sideways

Report Structure (Dashboard Sections)

1. Executive Summary — Top-level takeaways and recommended focus areas
2. Industry Trends — What's growing, declining, or emerging in the space
3. Competitor Landscape — Key players, funding stage, positioning
4. Existing Products — Products solving related problems, with feature summaries
5. Pain Points & Limitations — What users complain about (from social signals)
6. Market Gaps & Opportunities — Unserved niches or under-served segments
7. Recommended Next Steps — Actionable ideas for the founder to validate

Research Agent Flow

1. Query Decomposition — LLM breaks the founder's prompt into 3-5 parallel search queries
2. Parallel Search — Each query runs through DuckDuckGo independently
3. Result Aggregation — Deduplicate and rank results by relevance
4. Synthesis Pass — LLM analyzes aggregated data and produces structured report JSON
5. Response — Backend returns structured JSON; frontend renders interactive dashboard

Tech Stack Details

Backend (backend/)

• Framework: FastAPI with uvicorn
• Search: duckduckgo-search Python library (free, no API key)
• LLM Client: anthropic SDK with ANTHROPIC_BASE_URL override to Moonshot
• Data: In-memory storage (research sessions keyed by UUID)
• Key Files:
  • main.py — FastAPI app, CORS, routes
  • agent.py — Research orchestration (decompose → search → synthesize)
  • search.py — DuckDuckGo search wrapper with rate limiting
  • models.py — Pydantic models for request/response

Frontend (frontend/)

• Framework: React 18 + Vite
• Styling: Tailwind CSS (or plain CSS for speed)
• Key Components:
  • PromptInput — Natural language textarea with examples
  • ResearchDashboard — Report layout with collapsible sections
  • SectionCard — Reusable card for each report section
  • LoadingState — Progress indicator during research
• Key Files:
  • src/App.jsx — Main layout, routing
  • src/components/PromptInput.jsx
  • src/components/ResearchDashboard.jsx
  • src/api.js — Fetch wrapper for /research

Critical Design Decisions

1. DuckDuckGo over paid APIs — Zero API keys needed for demo. Can swap to Serper.dev/Tavily later.
2. No database — Research sessions stored in-memory. Simple for demo; add SQLite if history is needed.
3. Streaming response — FastAPI streams partial results so the dashboard populates incrementally (better UX for 30-60s wait).
4. One-shot vs. multi-turn — Strictly one-shot. No follow-up questions from the agent to keep it fast.

Verification Plan

1. Start backend: cd backend && uvicorn main:app --reload
2. Start frontend: cd frontend && npm run dev
3. Open http://localhost:5173, enter a prompt like:

   “I'm a solo developer with $50k budget, interested in B2B SaaS for the construction industry. I have React and Python skills.”

4. Verify dashboard renders with all 7 sections populated.
5. Verify research completes in under 60 seconds.

Open Questions

• Should research sessions be persisted to disk (SQLite) so reports survive server restarts?
• Should the dashboard allow exporting to PDF/Markdown?
• Should the agent suggest adjacent/nearby markets if the primary market looks saturated?