Product Blueprint

One person should be able to run a 24/7 AI marketing team. Reachly AI is not a scheduler — it is a B2C customer acquisition platform that discovers, understands, talks to and converts real people on the social accounts you already own.

Wedge: content tools sell reach; Reachly sells booked consultations and paid enrolments. Every feature is judged by whether it moves someone from "saw a post" to "paid".

• Beachhead: education & test-prep businesses in Canada (IELTS, English programs, tutoring, drama/enrichment) with 1–10 staff and CAD $200–2,000 offers.
• Adjacent wave 2: study-abroad and immigration consultants, coaches and consultants, online course sellers.
• Wave 3: creators, freelancers, local service businesses, indie brands.
• They already get DMs and comments they cannot answer fast enough, sell through conversation, and have no CRM. That is the pain we monetise.

"Turn my Instagram DMs and comments into booked consultations."

An IELTS school enters: Chinese international students in Canada, 15–18, IELTS need. Reachly listens to its own account's comments, DMs and mentions, detects IELTS intent, scores each person 0–100, drafts a personalised reply grounded in the school's own pricing and curriculum, waits for approval, sends, then follows up until the free speaking assessment is booked.

1. Sign up → business profile (industry, offer, price, region, language).
2. AI Persona wizard — 6 fields, 90 seconds.
3. Knowledge Base — paste website URL, upload brochure/pricing/FAQ. Auto-indexed.
4. Connect Instagram/Facebook via OAuth; capability matrix shown honestly.
5. Backfill: last 30 days of comments and DMs scored → instant "you have 7 warm leads".
6. Approve the first 5 AI replies (aha moment: it sounds like them, and it cites price).
7. Enable auto-reply for low-risk comment answers; DMs stay approval-gated.
8. First booked appointment → first sale → upgrade to Pro.

• Research Agent — builds the ICP, finds where the audience gathers, tracks competitor and trend signals. Output: audience hypotheses, content angles, keyword/hashtag watchlist.
• Listening Agent — polls/receives webhooks for authorised comments, DMs, mentions and engagement; classifies intent, urgency, language and spam.
• Lead Agent — scores Interest 30 / Need 25 / Budget 15 / Timing 15 / Engagement 15 → Cold <45, Warm 45–74, Hot ≥75, with reasons attached.
• Content Agent — posts, comments, DMs, follow-ups, short-video scripts, emails; personalised per lead from their own public content, never mail-merge.
• Sales Agent — consultative flow: goal → level → test date → target score → location → budget → pain point; one question per message; escalates to human on refunds, complaints, contracts or distress.
• CRM Agent — maintains the customer profile, stage, history, next action.
• Follow-up Agent — Day 1/3/7/14/30 cadence under frequency caps, quiet hours, opt-out and per-platform policy windows.
• Orchestrator — LangGraph state machine holding the shared lead state, tool permissions, approval gates and the learning write-back.

Target definition -> Listening -> Prospect discovery -> Profile ->
Intent scoring -> Personalised content -> Comment reply -> DM ->
Qualification -> CRM -> Follow-up -> Conversion -> Analytics

Inner loop, per signal:
Observe -> Understand -> Decide -> Recommend -> Approve -> Act -> Learn

Learn is the moat: every action stores {angle, opener length, question type, offer, time-of-day} against reply / booking / sale outcomes. Winning variants get higher sampling weight per segment, so the system compounds instead of plateauing.

• Frontend: React + TanStack Start (this app), SSR routes, TanStack Query.
• App backend: server functions / API routes on the edge for CRUD, OAuth callbacks, webhooks, Stripe.
• Agent runtime: Python FastAPI + LangGraph workers (long-running, retryable, observable) behind a queue.
• Queue & schedule: Redis + Celery for polling, follow-up timers, rate-limit buckets per platform per account.
• Data: PostgreSQL + pgvector for knowledge chunks and message embeddings; object storage for uploads.
• AI: gateway-routed LLMs — a strong model for sales reasoning, a cheap fast model for classification/scoring, embeddings for retrieval.
• Ops: PostHog product analytics, structured agent traces, Sentry, Stripe billing with credit metering.

organizations(id, name, plan, ai_credits, quiet_hours, timezone)
users(id, org_id, email, role)
personas(id, org_id, name, tone, languages[], target_customer,
         product, price, value_prop, disclose_ai, auto_low_risk)
knowledge_sources(id, org_id, type, uri, status)
knowledge_chunks(id, source_id, content, embedding vector(1536), meta)
social_accounts(id, org_id, platform, external_id, handle,
                scopes[], access_token_enc, refresh_token_enc,
                expires_at, status)
leads(id, org_id, platform, external_user_id, handle, name, locale,
      location, stage, intent, intent_confidence,
      score_interest, score_need, score_budget, score_timing,
      score_engagement, score_total, temperature,
      consent_basis, opted_out_at, last_contact_at, next_action_at)
conversations(id, lead_id, social_account_id, channel, platform_thread_id,
              category, summary, unread, policy_window_expires_at)
messages(id, conversation_id, direction, body, sent_at,
         platform_message_id, generated_by, tokens, sources_used[])
actions(id, org_id, lead_id, type, payload, risk, status,
        recommended_at, approved_by, executed_at, error)
followups(id, lead_id, sequence_step, scheduled_at, status)
outcomes(id, action_id, replied, booked, purchased, revenue,
         variant_tags[])
appointments(id, lead_id, starts_at, source, status)
usage_events(id, org_id, kind, units, cost_cents, created_at)
audit_log(id, org_id, actor, action, target, created_at, ip)

POST /api/oauth/:platform/start        -> consent redirect
GET  /api/oauth/:platform/callback     -> token exchange + encrypt
POST /api/public/webhooks/:platform    -> signature-verified events
GET  /api/inbox?filter=hot             -> triaged conversations
POST /api/conversations/:id/suggest    -> grounded draft reply
POST /api/actions/:id/approve|decline  -> human-in-the-loop gate
POST /api/leads/:id/score              -> rescore with reasons
POST /api/knowledge/sources            -> ingest + embed
POST /api/followups/run                -> cron, respects caps
POST /api/public/cron/listen           -> polling sweep
GET  /api/analytics/funnel             -> reach -> sale
POST /api/billing/checkout|portal      -> Stripe

• Per-platform OAuth 2.0 app with least-privilege scopes; PKCE where supported.
• Tokens encrypted at rest (envelope encryption, per-org key), refreshed by a worker before expiry, revoked instantly on disconnect.
• A capability registry per platform declares each action as api_automatable or recommend_only, plus rate limits and the messaging policy window (e.g. Instagram/Messenger 24h human-agent window, WhatsApp template rules). The orchestrator can only emit actions the registry allows.
• Webhook signature verification on every inbound event; replay protection.

• Layer 1 – Persona: identity, tone, languages, offer, price, disclosure rules.
• Layer 2 – Policy: platform + legal guardrails, quiet hours, frequency caps, escalation triggers. Non-overridable by user input.
• Layer 3 – Retrieval: top-k knowledge chunks with citations; hard rule that price/policy/outcome claims must come from retrieved text.
• Layer 4 – Lead context: profile, score breakdown, prior messages, objections, stage.
• Layer 5 – Task: classify / score / draft / follow-up, with structured JSON output ({reply, intent, confidence, score, next_action, requires_human, sources}).
• Prompt-injection defence: inbound social text is treated as untrusted data, never instructions.

Built-in, conversation-first CRM: stages New → Engaged → Qualified → Appointment → Won/Lost, each lead carrying platform identity, score breakdown with reasons, full cross-channel history, consent basis, opt-out flag, owner (agent or human) and one explicit next action with a due time. Nothing is a free-text note dump. HubSpot/Salesforce sync is a mapped outbound adapter, not a dependency.

Command Center answers three questions above the fold: what happened today (reached, conversations, qualified, appointments, sales, conversion rate), what needs me now (approval queue), and is it working (funnel + weekly trend). The AI Inbox is the daily driver: filter to 🔥 Hot, read the AI summary, press Approve & send. Leads & CRM is the pipeline view. Everything else is setup you touch once.

MVP platform: Instagram (with Messenger/Facebook Page as the same Meta connector). Reasons: the beachhead's audience (15–18 international students and their parents) lives there; Meta's Graph API officially supports reading comments, replying to comments, and DM conversations inside the 24-hour human-agent window; inbound DM volume already exists so we need no cold outreach to show value; one OAuth app covers two platforms.

MVP scope: registration & org, business profile, AI Persona, Knowledge Base ingest + retrieval, Instagram/Messenger OAuth, unified inbox with AI summary + intent, lead scoring, AI reply with human approval, built-in CRM, Day 1/3/7 follow-up, funnel analytics, Stripe subscription with AI credit metering.

Explicitly out of MVP: cold outreach, LinkedIn/TikTok/X/YouTube automation, autonomous sending, team seats, white label, marketplace.

W1  Foundations: auth, org model, schema, design system, Command Center shell
W2  Persona + Knowledge Base: ingest, chunk, embed, grounded answer test
W3  Meta OAuth + webhooks: token vault, capability registry, backfill 30 days
W4  Unified Inbox + classification, summary, intent scoring
W5  Lead scoring + CRM + approval queue (Observe->Act path complete end to end)
W6  Follow-up engine: cadence, quiet hours, caps, opt-out, appointment links
W7  Analytics funnel, outcome logging, learning loop v1, Stripe + credits
W8  Compliance review, rate-limit hardening, 10 design-partner onboardings, launch

Free (500 credits) → Pro $39 → Business $149 → Agency $499, plus $9 per extra 1,000 AI interactions. Blended AI cost ≈ $0.006 per interaction keeps Pro at ~78% and Business at ~84% gross margin. Target self-serve CAC $70–110 against ~$546 Pro LTV (14-month retention) ≈ 5x LTV:CAC. See the Pricing page for the full table.

• Dogfood: the product acquires its own customers on Instagram/LinkedIn — publish the receipts ("this AI booked 6 consultations this week").
• Vertical wedge content: "IELTS school growth" playbooks, free lead-score audit of a school's last 30 days of DMs.
• Design-partner cohort of 10 schools at $0 for 60 days in exchange for outcome data.
• Agency/consultant channel with revenue share; education-association partnerships.
• Free tools as SEO surface: ICP generator, DM reply grader, IELTS lead-score calculator.

• Row-level tenant isolation; least-privilege service roles; full audit log.
• Encrypted token vault, secrets never in the client, no PII in prompts beyond what the conversation requires.
• PIPEDA/GDPR: lawful basis recorded per lead, purpose limitation, retention windows, export and erasure endpoints, DPA and sub-processor list.
• CASL/CAN-SPAM style anti-spam: inbound-initiated or opted-in contact only, identity disclosed, one-click opt-out honoured permanently, frequency caps and quiet hours enforced in code, not in prompts.
• Platform policy: official APIs only, published rate limits respected, no scraping of protected data, no purchased personal data, no automation of actions a platform forbids, no impersonation of a human, AI identity disclosed when asked or required.
• Kill switch per org and per channel; every automated send is attributable and reversible in the log.

• Scale: per-account rate-limit buckets, queue partitioning by org, cheap-model routing for classification, embedding cache, read replicas, warm/cold conversation tiers, cost ceiling per org per day.
• Reliability: idempotent send keys so a retry never double-messages a real person; dead-letter review queue.
• Marketplace (post-PMF): publishable Agent Packs — persona + prompts + follow-up cadence + knowledge template for a vertical (IELTS school, immigration consultant, yoga studio), rated on real conversion data, revenue-shared with the author. Plus connector SDK and skill plugins (booking, payments, quizzes). The network effect is proven playbooks, not software.