Connecttr

A full-stack, multi-tenant SaaS that automates the entire B2B outbound-sales loop — from finding companies that actually need a product, to enriching contacts, scoring buying intent, sending personalized email, and learning from every reply to re-tune itself.

B2B outbound is manually expensive: finding the right companies, getting contact info, writing personalized emails, and knowing when to follow up. Connecttr closes that loop end-to-end for each client organization, from signal detection all the way to engagement analytics.

---

🏗️ Architecture & pipeline

React 19 (SPA)  ⇅ REST + SSE streaming ⇅  FastAPI (Python)  ⇅  MongoDB Atlas
       integrations:  Perplexity Sonar · Scrape.do · Jina.ai · Postmark · Google OAuth
       deploy:        Render (Gunicorn/UvicornWorker backend + static frontend)

The four phases

1. Lead discovery — Perplexity Sonar turns a client's website into a structured brief (ICP, buying signals, ready-made search queries); those queries hit a JS-rendering Google proxy; results are filtered and heuristically scored; contacts are enriched (Jina.ai → BeautifulSoup fallback) and upserted to MongoDB — streamed to the UI over SSE.
2. Intent scoring — a composite 0–100 score blending rule_score (0.30), llm_score (0.50, Perplexity), and engagement_score (0.20 from opens/clicks/replies).
3. Email generation & sending — personalized templates sent via Postmark; open/click/reply webhooks update engagement and recompute scores in real time.
4. Campaign health & learning — when observed vs. predicted conversion diverges by >5% (over 30+ events), the scoring weights (W_LLM / W_RULE) are auto-tuned in 1% increments, stored per org.

✨ Engineering highlights

• Adaptive scoring loop — the model's weights are not static; they self-adjust from real engagement outcomes per organization.
• Multi-tenancy & RBAC — every query is scoped by org_id; roles are owner / admin / member; destructive actions are RBAC-gated and written to an audit_logs collection.
• Streaming UX — long-running discovery is delivered incrementally via Server-Sent Events.
• Service auth — the internal /pipeline/* API is isolated behind a separate X-API-KEY header.
• Data model — users → organizations → campaigns → leads → emails (+ org_invites).

🗂️ Repository structure

├── front-end/          # React 19 SPA — Radix UI, Tailwind, framer-motion, Google OAuth
│   └── src/            # dashboard, auth context, protected routes, UI kit
├── back-end/
│   ├── app/
│   │   ├── api/        # auth, campaigns, email_hub, intent_analytics, orgs, pipeline
│   │   ├── services/   # company_analyzer, contact_enricher, intent, learning, health
│   │   ├── ml/         # scoring / prediction
│   │   ├── models/     # org, user, audit, invite, alert
│   │   └── scripts/    # index creation, SaaS migration, phase-3 simulation
│   └── *dry_run*.py    # audit & migration dry-run tooling with reports
└── docs/               # architecture overview + phased delivery plans

🧰 Tech stack

Frontend React 19 · Radix UI · Tailwind · framer-motion · Google OAuth · jwt-decode Backend FastAPI · MongoDB Atlas · Perplexity Sonar · Postmark · Scrape.do · Jina.ai Infra Render · Gunicorn/UvicornWorker · SSE streaming · API-key service auth

---

_{Author: Ayush Deo · MS CS @ USC · github.com/ayushdeo · Full-stack + applied LLM systems}