StartupAI

Inspiration

We’ve seen too many people with a strong idea get stuck on structure: they know what they want to build but struggle to turn it into a clear name, positioning, features, pricing, and pitch. Writing a one-pager or deck from scratch is slow and often ends up as a blank page. We wanted to flip that — one sentence in, full blueprint out — so anyone can go from “I have an idea” to “here’s the startup on a page” in seconds. We also wanted to support the way ideas often start: not only in text, but in screenshots, sketches, or mood boards. So we combined Reka for fast, structured generation (including vision for images) with Tavily for real-world context, giving founders a head start without losing the rigor of a real blueprint.

What it does

Startup in a Box turns a single sentence (e.g. “A meal-planning app for busy parents”) into a full, structured startup blueprint in seconds. Users get:

• Brand & positioning — Creative name, tagline, and category
• Target persona — Named persona with pain points and goals
• Core features — 5 feature cards with descriptions and icons
• Pricing model — Tiered pricing (e.g. Freemium/Subscription) with features per tier
• Tech stack — Frontend, backend, database, AI, devops, infrastructure, APIs, observability
• Launch roadmap — 8-week phased plan with concrete tasks
• Investor pitch — Problem, solution, market size, traction, ask, and unique advantage
• Scores — Confidence (1–100) and evaluation scores (solution, problem, market, etc.)

Users can also generate from an image (screenshot, sketch, mood board) via Reka Vision. Beyond generation, users can save blueprints, compare up to 3 side-by-side, export to PDF, browse a scoreboard of top ideas, and use competitor analysis (Tavily Search). An admin dashboard shows metrics (generation time, tokens, output length) and a leaderboard.

How we built it

• Frontend: React, TypeScript, Vite, React Router, shadcn-ui (Radix) + Tailwind CSS, React Hook Form, Zod, TanStack React Query, Recharts, next-themes for dark/light mode.
• Backend & data: Supabase — Postgres for blueprints, history, favorites, and metrics; Supabase Edge Functions (Deno) for all server-side logic.
• AI: Reka API (reka-flash) for text generation with strict JSON schemas and system prompts. Reka Vision (same API, multimodal) powers generate-startup-from-image: one image (screenshot, sketch, mood board) → full blueprint — Most Innovative Use of Reka's Vision API.
• Integrations: Tavily Search for real-time web search in competitor analysis (Best use of Tavily Search).
• Dev experience: ESLint, Vitest + React Testing Library, Supabase CLI for migrations and local function testing.

The flow: user enters an idea (or pastes an image URL) → frontend calls generate-startup or generate-startup-from-image → Reka returns JSON → we validate, optionally save to Supabase, and display the blueprint. History, compare, scoreboard, and dashboard all read from the same Supabase tables.

Challenges we ran into

• Stable JSON from the LLM — Getting Reka to return only valid JSON with no markdown or extra text was tricky. We fixed it with strict system instructions (“Respond ONLY with valid JSON”), clear schema examples in the user prompt, and client-side parsing with fallbacks for minor formatting issues.
• Balancing detail vs. speed — A full blueprint (name, persona, features, pricing, tech stack, roadmap, pitch, scores) is a lot of tokens. We tuned the schema to be rich but compact and used a single call with a well-structured prompt so one generation stays fast and under token limits.
• Optional integrations — Tavily is optional. We had to handle “no API key” gracefully so the app still works for generation and history when only Reka + Supabase are configured.
• Reka Vision from image — Designing a single prompt that works for any image (screenshot, wireframe, sketch) and still produces the full JSON blueprint required careful prompt design and robust JSON extraction.
• Edge Functions and CORS — Making sure the frontend (often on a different origin) could call Supabase Edge Functions without CORS errors required consistent headers and correct handling of preflight (OPTIONS) requests.

Accomplishments that we're proud of

• One idea → full blueprint in one shot — Name, tagline, persona, features, pricing, tech stack, 8-week roadmap, investor pitch, and confidence/evaluation scores from a single Reka call, with consistent JSON structure.
• Image → blueprint (Reka Vision) — Upload or paste an image URL (screenshot, sketch, mood board) and get a full startup blueprint in one call — a strong fit for Most Innovative Use of Reka's Vision API.
• Real-world usefulness — Outputs are specific and actionable (realistic tech stacks, concrete roadmap tasks, coherent pitch sections), not generic templates.
• Compare & scoreboard — Side-by-side comparison of up to 3 blueprints and a leaderboard by confidence score make it easy to evaluate and iterate on ideas.
• Competitor and market context — Tavily for live web search (Best use of Tavily Search) means blueprints can be grounded in real competitors and market signals.
• Polished UX — History with search/favorites, PDF export, dark mode, and an admin dashboard with metrics show we cared about the full product, not just the generator.

What we learned

• Prompt design is product design — The exact schema and wording in the Reka prompt directly shape output quality. Investing in a clear, explicit JSON schema and examples paid off for both text and vision.
• Edge Functions are a great fit for AI — Keeping API keys on the server (Supabase secrets), using Deno for the functions, and invoking from the frontend via Supabase client gave us a simple, secure backend without a separate Node server.
• Structured output beats free-form — Constraining the model to a fixed JSON structure improved reliability and made the frontend (and comparisons/export) much simpler than parsing prose.
• Optional integrations — Designing Tavily as optional from the start kept the core product working for everyone while allowing power users to add real-time competitor research.

What's next for Startup in a Box

• More AI providers — Support for other models (e.g. Claude, GPT) and optional “premium” models for higher-quality or longer blueprints.
• Templates and verticals — Pre-built templates (e.g. “B2B SaaS,” “Marketplace”) and vertical-specific prompts for more tailored outputs.
• Collaboration — Share blueprints with teammates, comment on sections, and iterate together.
• Deeper integrations — Richer use of Tavily (e.g. market size or trend snippets injected into the blueprint).
• Monetization — Freemium: N free generations per month, then paid tiers for unlimited generation, compare, export, and competitor analysis.

Built With

• react
• reka
• tailwind
• tavily
• typescript