package-lock.json ×

package.json ×

plan.md

# obrol.com — Tanya AI Bahasa Indonesia yang ngerti Konteks Lokal

> AI Q&A platform untuk Indonesian yang capek dapat jawaban ChatGPT "diterjemahkan asal" — obrol jawab pakai konteks BPJS, PPh, Kominfo, izin warung, sekolah negeri, dll, dengan sumber resmi yang clickable.

## §1 The Wedge

ChatGPT/Gemini/Claude lemah di pertanyaan spesifik-Indonesia. Coba:
- "Berapa biaya BPJS Kesehatan kelas 1 untuk 4 orang di 2026?"
- "Cara daftar SIUP UMKM online lewat OSS lewat HP?"
- "Anak saya tidak diterima zonasi PPDB SMP — apa yang bisa saya lakukan?"
- "PPh 21 karyawan gaji Rp 8jt tahun 2026 berapa potongan?"

Jawaban model umum: outdated (data 2022/2023), pakai bahasa stiff hasil terjemahan, sumber dari Wikipedia US, no actionable next step. Untuk pertanyaan Indonesian-context, hasilnya **lebih sering bikin user lebih bingung daripada lebih jelas**.

obrol.com narrow ke gap ini: **AI jawab pakai data + sumber Indonesia khusus**, dengan citation yang bisa diklik (bukan halusinasi). Wedge bukan "kita lebih pintar dari GPT" — wedge: **kita lebih lokal**.

Target user pertama: ortu murid SD/SMP/SMA, owner UMKM yang nggak ada admin, lansia pensiunan yang baru pertama urus BPJS sendiri. Mereka tidak peduli model apa di belakang — mereka peduli jawabannya benar dan bisa langsung dipakai.

## §2 Money Stack

**Tier Gratis** (akuisisi engine):
- Unlimited tanya
- Jawaban single-source (no chain reasoning)
- 24 jam history
- Indonesian-grounded knowledge base auto-refresh tiap minggu

**Tier Pendamping · Rp 19.000/bulan** (atau Rp 159k/tahun):
- Unlimited history & saved threads
- Multi-source citation per jawaban (sampai 5 sumber)
- Voice input via mic (Indonesian STT) — penting buat ortu/UMKM yang ngetik lambat
- "Pakar Mode" — jawaban lebih dalam dengan reasoning chain (slower, 8-15 detik)
- Document upload (max 10 PDF/bulan) untuk tanya "isi kontrak ini boleh nggak begini?"
- Family sharing — 4 anggota dalam satu langganan

**Tier UKM · Rp 99.000/bulan** (B2B kecil):
- Semua di Pendamping
- "Kustomisasi konteks bisnis" — upload regulasi spesifik industri (apotek, restoran, salon) sekali, jawaban grounded ke konteks itu seterusnya
- WhatsApp Business bot integration (chat AI dari WA langsung, jawaban auto-cite)
- Multi-staff seat (5 user di satu akun)

**Affiliate compound layer:**
- User tanya "rekomendasi asuransi kesehatan keluarga selain BPJS" → suggest produk Allianz/AXA via referral (komisi 1-3% premium)
- "Cara invest reksadana pertama kali" → Bibit/Bareksa referral
- Konservatif: 1 dari 50 jawaban convert affiliate → Rp 5-25rb tambahan per active user/bulan

**Target ekonomi tahun 1:**
- 80,000 free users (akuisisi SEO + viral)
- 3% convert Pendamping = 2,400 × Rp 19rb = Rp 45,6jt/bulan
- 100 UKM = Rp 9,9jt/bulan
- Affiliate revenue ~Rp 30jt/bulan
- **Total MRR target: Rp 85jt/bulan** (~$5,500)

## §3 Acquisition triple-stack

1. **Programmatic SEO** — generate 5,000+ canonical "Cara X di Y 2026" pages dari Indonesian question patterns. Each page rendered by obrol's AI lalu reviewed sekilas, dipublish sebagai SEO landing. Long-tail traffic compound rapidly (lihat pattern langbuana.com & ourfuntech.com di portofolio).
2. **TikTok edutainment** — "5 hak ortu murid kalau anak gak diterima zonasi" type content, link → obrol.com untuk tanya kasus spesifik. Indonesian audience LAPAR sama actionable bureaucracy hacks.
3. **WhatsApp community widget** — tombol "Ask obrol" yang bisa di-embed di forum/group/komunitas. Klik tombol → buka link obrol.com dengan pertanyaan pre-filled.

## §4 AI automation map

| Loop | Model | Kenapa |
|---|---|---|
| Free-tier general Q&A (60% traffic) | Gemini 2.0 Flash Lite | Cepat + super cheap, Indonesian baik |
| Pakar Mode reasoning (Pendamping) | Gemini 2.5 Pro | Reasoning chain panjang, sumber multipel |
| RAG retrieval atas Indonesian gov + news sites | Workers AI embeddings + Vectorize | Free, edge-local |
| Voice STT Indonesian | ElevenLabs Scribe | Best Indonesian STT accuracy 2026 |
| TTS jawaban (Premium opt-in) | ElevenLabs v3 Bahasa | Natural Indonesian voice |
| Citation extractor + URL validator | rule-based + Workers fetch | Pastikan link tidak dead |
| SEO content generator (5000+ Q pages) | Gemini Flash batch | Murah, scale |

**Cost ceiling per active user:**
- Free: ~Rp 50/user/bulan (40 query × Gemini Flash Lite)
- Pendamping: ~Rp 800/user/bulan (mix Flash + Pro)
- Margin healthy

## §5 Cloudflare architecture (100% serverless)

- **Worker:** `obrol-web` — chat UI + API orchestrator
- **D1:** `obrol_db` — users, threads, citations cache, saved questions
- **KV:** `obrol-cache` — popular Q answer cache (24h TTL — top 5000 questions skip AI entirely, served from edge)
- **Vectorize:** `obrol-knowledge` — Indonesian gov regulations + news sites embeddings (refreshed weekly via cron)
- **R2:** `obrol-docs` — user-uploaded PDFs (Premium, encrypted at rest)
- **Workers AI:** bge embeddings + Llama-3 fallback if Gemini quota habis
- **Queue:** `obrol-batch-gen` — SEO content generation batch
- **Durable Object:** rate limit per user (cheap + atomic)

## §6 90-day execution

**Week 1-2 — Core Q&A loop**
- Landing dengan search box besar, AI jawab pakai Gemini Flash + simple citation
- D1 schema + per-IP rate limit (10/hari free tier)
- Top 200 hot Indonesian queries cached di KV

**Week 3-4 — Knowledge grounding (RAG)**
- Scrape + embed: kemkes.go.id, kemdikbud.go.id, bpjs-kesehatan.go.id, pajak.go.id, oss.go.id, kominfo.go.id, kompas.com, detik.com (15-30 sumber tier-1)
- Cron weekly refresh
- Citation footnotes clickable di jawaban

**Week 5-6 — Premium tier (Pendamping)**
- Mayar QRIS Rp 19rb subscription
- Voice input (ElevenLabs Scribe API)
- Pakar Mode (Gemini 2.5 Pro)
- Saved threads + history
- Document upload + tanya isi PDF

**Week 7-8 — SEO programmatic launch**
- Generate 500 pillar Q pages dari batch keyword research
- Pre-render HTML + JSON-LD FAQ schema per page
- Submit sitemap to GSC + IndexNow

**Week 9-12 — UKM tier + WA bot**
- WA Business API integration
- Multi-seat sub-accounts
- Konteks-bisnis upload (apoteker, restaurant, salon templates)
- Affiliate suggestion layer (1-of-50 jawaban suggest produk relevant)

## §7 Moat

**Indonesian-grounded knowledge base** = updated weekly dari sumber resmi. Untuk competitor mau replikasi, butuh: scraping infra + embedding pipeline + curation + Indonesian context awareness. Bukan rocket science, tapi consistent maintain 1+ tahun = real switching cost.

**SEO content compound** = 5000 long-tail Indonesian Q pages = traffic snowball yang ChatGPT/Gemini tidak bisa kalahkan di Google (mereka tidak punya canonical Indonesian pages indexable).

**Voice-first option** = significant moat untuk demografi ortu/UMKM yang ngetik lambat dan tidak mau bayar Rp 300rb ChatGPT Plus. Indonesian STT yang akurat itu valuable barrier.

## §8 Mandatory standards

- 100% Cloudflare worker-cf (no PM2)
- Citation MUST link to real Indonesian source — dead-link detection via cron
- Anonim by default — sign-up cuma diperlukan untuk save thread (Premium)
- AI responses MUST flag uncertainty ("Saya tidak yakin sumber terbaru, silakan cek ke [link]")
- User data encrypted at rest in D1 + R2
- Comply UU PDP — user delete request → full data wipe dalam 24 jam
- Indonesian-language UI only (untuk fokus + branding clarity)

## §9 Competitor synthesis

- **ChatGPT (OpenAI):** stale Indonesian knowledge cutoff, no source citation, $20/mo prohibitive
- **Gemini (Google):** better Indonesian than GPT, but no grounded citation, sometimes refuse Indonesian-specific Qs
- **Perplexity:** has citation but rarely cites Indonesian sources, English UX
- **Cermati / Lifepal:** content-portal style, slow, ads-cluttered, no AI Q&A
- **DetikEdu / KompasTanya:** content portals, no AI interaction
- **Quora Indonesia:** dead — moderation collapsed in 2022, full of spam

**Gap:** chat-first + Indonesian-grounded + voice option + Rp 19rb tier point. Belum ada.

## §10 Missing parts audit

- [ ] Vectorize index untuk Indonesian gov sites — scope mana yang prioritas (BPJS, OSS, Kominfo, Kemdikbud, Pajak)
- [ ] Mayar account untuk QRIS subscription
- [ ] ElevenLabs Indonesian voice clone test (untuk TTS Premium)
- [ ] OpenRouter / Gemini API key budget alarm
- [ ] Privacy policy + ToS reviewed UU PDP-style
- [ ] WA Business API — apakah pakai japri.com infra existing (shared WA `+6285353182303`) atau dapat number terpisah
- [ ] Brand mascot — ide: "Si Obrol" — character ramah seperti owl/kakek bijak yang nyantai (vs Si Duit yang ramah-Jakarta)
- [ ] SEO keyword research — beli list Indonesian long-tail Q dari Ahrefs / Semrush untuk awal
- [ ] BAPPEBTI/OJK consultation — apakah affiliate finansial (BPJS/asuransi suggestion) butuh license terdaftar

---

## ⚙ HARD CONSTRAINTS (enforced for all sites)

This domain MUST operate within these constraints — no exceptions:

- **100% Cloudflare serverless** — Workers + D1 + R2 + KV + Workers AI + Vectorize. NEVER PM2, NEVER VPS, NEVER Docker in production path.
- **100% AI-automated** — every customer interaction, every moderation decision, every transaction reconcile = AI. No manual queue, no live human chat support, no physical fulfillment.
- **1-operator solo** — one person can run the entire operation from a phone. No team meetings, no shared inbox, no shift rotation.
- **WhatsApp AI bot** for all support (24/7, instant response, no SLA promises that need humans).
- **Mayar QRIS** for all Indonesian payments (subscription auto-renew, no manual invoicing).
- **Indonesian UI primary** — bahasa-first, English fallback only where unavoidable.
- **Privacy** — opt-in only, delete-on-request honored within 24h (cron-driven).
- **No physical goods, no inventory** — digital products + affiliate referrals only.

**If the plan above describes any flow that violates these constraints, treat the plan as ASPIRATIONAL only and rework before building. The constraint trifecta wins.**

PREVIEW

obrol.com — Tanya AI Bahasa Indonesia yang ngerti Konteks Lokal

AI Q&A platform untuk Indonesian yang capek dapat jawaban ChatGPT "diterjemahkan asal" — obrol jawab pakai konteks BPJS, PPh, Kominfo, izin warung, sekolah negeri, dll, dengan sumber resmi yang clickable.

§1 The Wedge

ChatGPT/Gemini/Claude lemah di pertanyaan spesifik-Indonesia. Coba:

"Berapa biaya BPJS Kesehatan kelas 1 untuk 4 orang di 2026?"
"Cara daftar SIUP UMKM online lewat OSS lewat HP?"
"Anak saya tidak diterima zonasi PPDB SMP — apa yang bisa saya lakukan?"
"PPh 21 karyawan gaji Rp 8jt tahun 2026 berapa potongan?"

Jawaban model umum: outdated (data 2022/2023), pakai bahasa stiff hasil terjemahan, sumber dari Wikipedia US, no actionable next step. Untuk pertanyaan Indonesian-context, hasilnya lebih sering bikin user lebih bingung daripada lebih jelas.

obrol.com narrow ke gap ini: AI jawab pakai data + sumber Indonesia khusus, dengan citation yang bisa diklik (bukan halusinasi). Wedge bukan "kita lebih pintar dari GPT" — wedge: kita lebih lokal.

§2 Money Stack

Tier Gratis (akuisisi engine):

Unlimited tanya
Jawaban single-source (no chain reasoning)
24 jam history
Indonesian-grounded knowledge base auto-refresh tiap minggu

Tier Pendamping · Rp 19.000/bulan (atau Rp 159k/tahun):

Unlimited history & saved threads
Multi-source citation per jawaban (sampai 5 sumber)
Voice input via mic (Indonesian STT) — penting buat ortu/UMKM yang ngetik lambat
"Pakar Mode" — jawaban lebih dalam dengan reasoning chain (slower, 8-15 detik)
Document upload (max 10 PDF/bulan) untuk tanya "isi kontrak ini boleh nggak begini?"
Family sharing — 4 anggota dalam satu langganan

Tier UKM · Rp 99.000/bulan (B2B kecil):

Semua di Pendamping
"Kustomisasi konteks bisnis" — upload regulasi spesifik industri (apotek, restoran, salon) sekali, jawaban grounded ke konteks itu seterusnya
WhatsApp Business bot integration (chat AI dari WA langsung, jawaban auto-cite)
Multi-staff seat (5 user di satu akun)

Affiliate compound layer:

User tanya "rekomendasi asuransi kesehatan keluarga selain BPJS" → suggest produk Allianz/AXA via referral (komisi 1-3% premium)
"Cara invest reksadana pertama kali" → Bibit/Bareksa referral
Konservatif: 1 dari 50 jawaban convert affiliate → Rp 5-25rb tambahan per active user/bulan

Target ekonomi tahun 1:

80,000 free users (akuisisi SEO + viral)
3% convert Pendamping = 2,400 × Rp 19rb = Rp 45,6jt/bulan
100 UKM = Rp 9,9jt/bulan
Affiliate revenue ~Rp 30jt/bulan
Total MRR target: Rp 85jt/bulan (~$5,500)

§3 Acquisition triple-stack

Programmatic SEO — generate 5,000+ canonical "Cara X di Y 2026" pages dari Indonesian question patterns. Each page rendered by obrol's AI lalu reviewed sekilas, dipublish sebagai SEO landing. Long-tail traffic compound rapidly (lihat pattern langbuana.com & ourfuntech.com di portofolio).
TikTok edutainment — "5 hak ortu murid kalau anak gak diterima zonasi" type content, link → obrol.com untuk tanya kasus spesifik. Indonesian audience LAPAR sama actionable bureaucracy hacks.
WhatsApp community widget — tombol "Ask obrol" yang bisa di-embed di forum/group/komunitas. Klik tombol → buka link obrol.com dengan pertanyaan pre-filled.

§4 AI automation map

Loop	Model	Kenapa
Free-tier general Q&A (60% traffic)	Gemini 2.0 Flash Lite	Cepat + super cheap, Indonesian baik
Pakar Mode reasoning (Pendamping)	Gemini 2.5 Pro	Reasoning chain panjang, sumber multipel
RAG retrieval atas Indonesian gov + news sites	Workers AI embeddings + Vectorize	Free, edge-local
Voice STT Indonesian	ElevenLabs Scribe	Best Indonesian STT accuracy 2026
TTS jawaban (Premium opt-in)	ElevenLabs v3 Bahasa	Natural Indonesian voice
Citation extractor + URL validator	rule-based + Workers fetch	Pastikan link tidak dead
SEO content generator (5000+ Q pages)	Gemini Flash batch	Murah, scale

Cost ceiling per active user:

Free: ~Rp 50/user/bulan (40 query × Gemini Flash Lite)
Pendamping: ~Rp 800/user/bulan (mix Flash + Pro)
Margin healthy

§5 Cloudflare architecture (100% serverless)

Worker: obrol-web — chat UI + API orchestrator
D1: obrol_db — users, threads, citations cache, saved questions
KV: obrol-cache — popular Q answer cache (24h TTL — top 5000 questions skip AI entirely, served from edge)
Vectorize: obrol-knowledge — Indonesian gov regulations + news sites embeddings (refreshed weekly via cron)
R2: obrol-docs — user-uploaded PDFs (Premium, encrypted at rest)
Workers AI: bge embeddings + Llama-3 fallback if Gemini quota habis
Queue: obrol-batch-gen — SEO content generation batch
Durable Object: rate limit per user (cheap + atomic)

§6 90-day execution

Week 1-2 — Core Q&A loop

Landing dengan search box besar, AI jawab pakai Gemini Flash + simple citation
D1 schema + per-IP rate limit (10/hari free tier)
Top 200 hot Indonesian queries cached di KV

Week 3-4 — Knowledge grounding (RAG)

Scrape + embed: kemkes.go.id, kemdikbud.go.id, bpjs-kesehatan.go.id, pajak.go.id, oss.go.id, kominfo.go.id, kompas.com, detik.com (15-30 sumber tier-1)
Cron weekly refresh
Citation footnotes clickable di jawaban

Week 5-6 — Premium tier (Pendamping)

Mayar QRIS Rp 19rb subscription
Voice input (ElevenLabs Scribe API)
Pakar Mode (Gemini 2.5 Pro)
Saved threads + history
Document upload + tanya isi PDF

Week 7-8 — SEO programmatic launch

Generate 500 pillar Q pages dari batch keyword research
Pre-render HTML + JSON-LD FAQ schema per page
Submit sitemap to GSC + IndexNow

Week 9-12 — UKM tier + WA bot

WA Business API integration
Multi-seat sub-accounts
Konteks-bisnis upload (apoteker, restaurant, salon templates)
Affiliate suggestion layer (1-of-50 jawaban suggest produk relevant)

§7 Moat

Indonesian-grounded knowledge base = updated weekly dari sumber resmi. Untuk competitor mau replikasi, butuh: scraping infra + embedding pipeline + curation + Indonesian context awareness. Bukan rocket science, tapi consistent maintain 1+ tahun = real switching cost.

SEO content compound = 5000 long-tail Indonesian Q pages = traffic snowball yang ChatGPT/Gemini tidak bisa kalahkan di Google (mereka tidak punya canonical Indonesian pages indexable).

Voice-first option = significant moat untuk demografi ortu/UMKM yang ngetik lambat dan tidak mau bayar Rp 300rb ChatGPT Plus. Indonesian STT yang akurat itu valuable barrier.

§8 Mandatory standards

100% Cloudflare worker-cf (no PM2)
Citation MUST link to real Indonesian source — dead-link detection via cron
Anonim by default — sign-up cuma diperlukan untuk save thread (Premium)
AI responses MUST flag uncertainty ("Saya tidak yakin sumber terbaru, silakan cek ke [link]")
User data encrypted at rest in D1 + R2
Comply UU PDP — user delete request → full data wipe dalam 24 jam
Indonesian-language UI only (untuk fokus + branding clarity)

§9 Competitor synthesis

ChatGPT (OpenAI): stale Indonesian knowledge cutoff, no source citation, $20/mo prohibitive
Gemini (Google): better Indonesian than GPT, but no grounded citation, sometimes refuse Indonesian-specific Qs
Perplexity: has citation but rarely cites Indonesian sources, English UX
Cermati / Lifepal: content-portal style, slow, ads-cluttered, no AI Q&A
DetikEdu / KompasTanya: content portals, no AI interaction
Quora Indonesia: dead — moderation collapsed in 2022, full of spam

Gap: chat-first + Indonesian-grounded + voice option + Rp 19rb tier point. Belum ada.

§10 Missing parts audit

Vectorize index untuk Indonesian gov sites — scope mana yang prioritas (BPJS, OSS, Kominfo, Kemdikbud, Pajak)
Mayar account untuk QRIS subscription
ElevenLabs Indonesian voice clone test (untuk TTS Premium)
OpenRouter / Gemini API key budget alarm
Privacy policy + ToS reviewed UU PDP-style
WA Business API — apakah pakai japri.com infra existing (shared WA +6285353182303) atau dapat number terpisah
Brand mascot — ide: "Si Obrol" — character ramah seperti owl/kakek bijak yang nyantai (vs Si Duit yang ramah-Jakarta)
SEO keyword research — beli list Indonesian long-tail Q dari Ahrefs / Semrush untuk awal
BAPPEBTI/OJK consultation — apakah affiliate finansial (BPJS/asuransi suggestion) butuh license terdaftar

⚙ HARD CONSTRAINTS (enforced for all sites)

This domain MUST operate within these constraints — no exceptions:

100% Cloudflare serverless — Workers + D1 + R2 + KV + Workers AI + Vectorize. NEVER PM2, NEVER VPS, NEVER Docker in production path.
100% AI-automated — every customer interaction, every moderation decision, every transaction reconcile = AI. No manual queue, no live human chat support, no physical fulfillment.
1-operator solo — one person can run the entire operation from a phone. No team meetings, no shared inbox, no shift rotation.
WhatsApp AI bot for all support (24/7, instant response, no SLA promises that need humans).
Mayar QRIS for all Indonesian payments (subscription auto-renew, no manual invoicing).
Indonesian UI primary — bahasa-first, English fallback only where unavoidable.
Privacy — opt-in only, delete-on-request honored within 24h (cron-driven).
No physical goods, no inventory — digital products + affiliate referrals only.

If the plan above describes any flow that violates these constraints, treat the plan as ASPIRATIONAL only and rework before building. The constraint trifecta wins.

AI ASSISTANT

Ask AI to research, improve, or generate content.

Try: "Research competitors for this niche"

Actions