LaunchLens

Turn customer noise into insight.

LaunchLens is an AI research workflow for marketing and product teams. Feed it an audience and a question — it scrapes the voice of the customer, clusters it into structured insights, runs probing AI customer interviews, and synthesizes campaign-ready positioning and messaging.

Stack: Next.js 15 (App Router) · Tailwind 4 · Supabase (auth + Postgres + RLS) · OpenAI API · Vercel. No LangChain, no vector DB, no Docker.

---

What it does

Three agents, each replacing a slow piece of the research stack:

1. Insight Engine — scrapes Reddit, TikTok (via Apify), and the open web; OpenAI clusters raw voice into four structured insight types with verbatim quotes:

   • belief — what people think is true
   • goal — what they actually want
   • context — when and where they use the category
   • pattern — repeated market tensions

2. AI Customer Interviews — a grounded chat interviewer that asks one question at a time, probes emotional drivers, and listens for contradictions. Grounded in whichever insights the Insight Engine has already pulled.

3. Report Generator — synthesizes insights + interview takeaways into bold positioning reframes (from → to) and campaign-ready messaging. Exports to Markdown today; Notion/slides are a TODO.

---

Project structure

launch-lens/
├── src/
│   ├── app/
│   │   ├── page.tsx                      — landing page
│   │   ├── (auth)/
│   │   │   ├── login/                    — login page
│   │   │   ├── signup/                   — signup page
│   │   │   ├── AuthForm.tsx
│   │   │   └── actions.ts                — login / signup / logout server actions
│   │   ├── dashboard/
│   │   │   ├── page.tsx                  — project list
│   │   │   └── new/
│   │   │       ├── page.tsx              — new project form
│   │   │       └── actions.ts            — createProject server action
│   │   ├── projects/[id]/
│   │   │   ├── layout.tsx                — shared header + tabs
│   │   │   ├── page.tsx                  — Insights tab
│   │   │   ├── interviews/page.tsx       — Interviews tab
│   │   │   └── report/                   — Report tab (+ export)
│   │   ├── api/projects/[id]/
│   │   │   ├── research/route.ts         — POST: scrape + extract (NDJSON stream)
│   │   │   ├── interview/route.ts        — POST: start | turn (text stream)
│   │   │   └── report/route.ts           — POST: generate positioning + messaging
│   │   ├── globals.css                   — Tailwind 4 + theme tokens
│   │   └── layout.tsx
│   ├── components/
│   │   ├── InsightCard.tsx
│   │   ├── InsightsGrid.tsx
│   │   ├── InterviewChat.tsx             — streaming chat UI
│   │   ├── ResearchRunner.tsx            — live scraping progress
│   │   ├── ProjectTabs.tsx
│   │   ├── Nav.tsx, Logo.tsx
│   ├── lib/
│   │   ├── ai/                           — OpenAI SDK + JSON / streaming helpers
│   │   ├── prompts.ts                    — all agent prompts
│   │   ├── scrapers/
│   │   │   ├── index.ts                  — parallel runner w/ progress events
│   │   │   ├── reddit.ts                 — reddit.com JSON + top comments
│   │   │   ├── web.ts                    — DDG search + readable extraction
│   │   │   ├── tiktok.ts                 — Apify (optional via APIFY_TOKEN)
│   │   │   └── types.ts
│   │   ├── supabase/
│   │   │   ├── client.ts                 — browser client
│   │   │   ├── server.ts                 — server + service-role client
│   │   │   └── middleware.ts             — session refresh + route gating
│   │   └── types.ts
│   └── middleware.ts
├── supabase/
│   ├── migrations/0001_init.sql          — tables + RLS
│   └── seed.sql                          — demo "Reusable water bottles for commuters"
├── next.config.ts
├── postcss.config.mjs
├── tsconfig.json
├── package.json
└── .env.example

---

Setup

1. Install

npm install
cp .env.example .env.local

Fill in .env.local:

var	what
NEXT_PUBLIC_SUPABASE_URL	from your Supabase project
NEXT_PUBLIC_SUPABASE_ANON_KEY	from your Supabase project
SUPABASE_SERVICE_ROLE_KEY	from your Supabase project (server only)
OPENAI_API_KEY	from https://platform.openai.com/api-keys
APIFY_TOKEN	optional — enables TikTok scraping
REDDIT_USER_AGENT	optional — override UA used on Reddit's public JSON

2. Supabase

Create a new Supabase project, then apply the migration. You have two options:

Option A — dashboard: Open supabase/migrations/0001_init.sql and paste it into the SQL editor in your Supabase dashboard.

Option B — CLI:

supabase link --project-ref <your-ref>
supabase db push

Auth: in Authentication → Providers, confirm Email is enabled. If you want instant sign-ups without email confirmation for local dev, set Authentication → Settings → Confirm email to off.

3. Run

npm run dev
# → http://localhost:3000

4. Demo seed (optional)

After creating at least one user, grab the auth.users.id and seed the "Reusable water bottles for commuters" demo project:

psql "$SUPABASE_DB_URL" -v user_id="'<your-uuid>'" -f supabase/seed.sql

---

Deployment (Vercel)

1. Push this repo to GitHub.
2. Import it in Vercel.
3. Add the same env vars as .env.local to Settings → Environment Variables.
4. Set Vercel's Function Max Duration to 300s on Pro (the research route can take a minute or two when sources are slow). The route already sets export const maxDuration = 300;.

No build config overrides needed — npm run build is the default.

---

Data model

All tables live under the public schema. RLS is on, scoped to the owning user_id via projects.

• projects — one per research workflow. Status: draft | running | ready | error.
• research_sources — raw scraped items (reddit posts/comments, web pages, tiktok captions).
• insights — extracted insights. type ∈ {belief, goal, context, pattern}, with confidence.
• quotes — verbatim supporting quotes per insight, with source URLs.
• interviews — AI customer interviews. Transcript is stored as jsonb.
• reports — generated positioning + messaging (jsonb).

See supabase/migrations/0001_init.sql.

---

How the agents talk to OpenAI

All prompts live in src/lib/prompts.ts and return strict JSON (parsed defensively by the helpers in src/lib/ai).

Models used (defaults, overridable via env):

• Insight extraction, interview turns, interview summaries — gpt-4.1-mini.
• Final report synthesis — gpt-4.1.

Prompt design principles baked in:

• Prefer specific over generic. The insight extractor is explicitly told that "users care about health" is a failure mode.
• Real human language only. Quotes must be verbatim with a source_index pointer into the supplied sources.
• Surface tensions. Contradictions and from → to reframes are first-class.
• One question at a time. The interviewer is instructed to probe, not lecture.

No proprietary research framework is exposed in the UI or prompts — we use "Customer Beliefs / Customer Goals / Usage Contexts / Market Patterns".

---

How research progress streams

POST /api/projects/:id/research returns an NDJSON stream — one JSON event per line:

{"type":"started","source":"reddit"}
{"type":"source_done","source":"reddit","count":34}
{"type":"stage","stage":"extract","message":"synthesizing 34 items with OpenAI"}
{"type":"insights_ready","count":9}
{"type":"done"}

The client reads this with a plain fetch + ReadableStream. See src/components/ResearchRunner.tsx.

The interview endpoint streams raw text deltas (no NDJSON framing needed — it's a single continuous assistant message).

---

What's stubbed

• TikTok scraping requires APIFY_TOKEN. If unset, the TikTok source is skipped cleanly and the rest of the pipeline still runs.
• Slides export isn't implemented. Markdown export works. Notion export is a single POST to /v1/pages away if you want to wire it up.
• Interview summarization into new insights — the prompt supports it, but we don't auto-promote interview findings into the insights table (you currently re-run research to regenerate). Easy extension.

---

Feedback / contributing

PRs welcome. Keep it simple, fast, hackable.