Technical Architecture

System Overview

Ping CRM is a three-tier application: a Next.js frontend communicates with a FastAPI backend, which persists data in PostgreSQL and uses Redis for caching, task brokering, and ephemeral state. Celery workers handle background sync, scoring, and AI-powered suggestion generation.

┌─────────────────┐       ┌──────────────────────┐       ┌────────────────┐
│   Next.js 15    │──────>│   FastAPI (async)     │──────>│  PostgreSQL    │
│   React 19      │ REST  │   Pydantic schemas    │  ORM  │  UUID PKs      │
│   Tailwind v4   │<──────│   JWT auth            │<──────│  GIN indexes   │
└─────────────────┘       └──────────┬───────────┘       └────────────────┘
                                     │
                          ┌──────────▼───────────┐
                          │   Redis              │
                          │   - Celery broker    │
                          │   - Rate gates       │
                          │   - Tweet cache      │
                          │   - OAuth nonces     │
                          └──────────┬───────────┘
                                     │
                          ┌──────────▼───────────┐
                          │   Celery Workers     │
                          │   - Gmail sync       │
                          │   - Telegram sync    │
                          │   - Twitter polling  │
                          │   - Score refresh    │
                          │   - AI suggestions   │
                          └──────────────────────┘

External services: Gmail API (OAuth), Telegram MTProto (Telethon), Twitter/X via Bird CLI (primary, cookie-based) with Twitter API v2 (OAuth 2.0 PKCE) as fallback, Anthropic Claude API (event classification and message composition), and a Chrome extension for LinkedIn profile data ingestion.

Backend Architecture

The FastAPI application lives under backend/app/ and follows a layered structure:

backend/app/
├── main.py              # FastAPI app factory, CORS, router mounting
├── api/                 # Route handlers (REST endpoints)
│   ├── contacts.py      # CRUD, CSV import, merge, search
│   ├── interactions.py  # Interaction timeline
│   ├── suggestions.py   # Follow-up suggestions, snooze, dismiss
│   ├── identity.py      # Identity resolution scan, review, merge
│   ├── organizations.py # Company records and stats
│   ├── notifications.py # In-app notification feed
│   ├── auth.py          # Login, register, OAuth callbacks
│   ├── telegram.py      # Connect, sync, send messages
│   ├── twitter.py       # OAuth PKCE flow, polling
│   └── settings.py      # User preferences
├── models/              # SQLAlchemy 2.x async ORM models
├── schemas/             # Pydantic request/response schemas
├── services/            # Business logic layer
├── integrations/        # Third-party API clients
└── core/                # Config, auth, database, encryption, Redis

API Response Envelope

All endpoints return a typed generic envelope. This provides a consistent contract for the frontend:

from pydantic import BaseModel, Generic, TypeVar

T = TypeVar("T")

class Envelope(BaseModel, Generic[T]):
    data: T | None = None
    error: str | None = None
    meta: dict | None = None

Endpoints declare response_model=Envelope[SomePayloadType], giving full type safety through the stack. For example:

@router.get("/contacts/{contact_id}", response_model=Envelope[ContactRead])
async def get_contact(contact_id: uuid.UUID, db: AsyncSession = Depends(get_db)):
    ...
    return Envelope(data=contact)

Services Layer

Service	Responsibility
`followup_engine`	Two-pool suggestion generator (Pool A: active relationships, Pool B: dormant revival)
`identity_resolution`	Tier 1 deterministic + Tier 2 probabilistic duplicate detection and merge
`scoring`	Relationship score calculation (0-10 scale)
`event_classifier`	LLM-based classification of tweets and bio changes into structured events
`message_composer`	Context-aware AI draft generation with tone matching
`auto_tagger`	LLM-driven contact tagging with taxonomy management
`organization_service`	Company grouping, domain extraction, stats materialized view
`contact_search`	Full-text search across contacts
`telegram_service`	Higher-level Telegram orchestration with caching
`bio_refresh`	Twitter/Telegram bio change detection and tracking

Integrations

Integration	Protocol	Auth
Gmail	Gmail API (REST)	OAuth 2.0 (per-account tokens via `GoogleAccount` model)
Telegram	MTProto via Telethon	Phone number + session string (encrypted at rest)
Twitter/X	Bird CLI (primary) + X API v2 (fallback)	Cookie-based (`AUTH_TOKEN` + `CT0`) / OAuth 2.0 PKCE
Google Calendar	Calendar API (REST)	Shared OAuth tokens with Gmail
LinkedIn	Chrome extension (browser-side scrape)	Extension ID (`CHROME_EXTENSION_ID`); no LinkedIn API key required

Database Schema

PostgreSQL with UUID primary keys throughout. Key models and their relationships:

┌──────────┐     ┌───────────┐     ┌──────────────┐
│   User   │────<│  Contact   │────<│  Interaction  │
│          │     │            │     │  (email/tg/x) │
│          │     │ org_id ──────>┌──┴──────────────┘
│          │     └─────┬──────┘ │
│          │           │        │  ┌──────────────────┐
│          │           ├───────<│  DetectedEvent      │
│          │           │        │  (job_change, etc.) │
│          │           │        └──────────────────────┘
│          │     ┌─────▼──────┐
│          │────<│FollowUp    │
│          │     │Suggestion  │
│          │     └────────────┘
│          │
│          │────<│ Notification │
│          │────<│ GoogleAccount │
└──────────┘
           ┌────────────────┐
           │ IdentityMatch  │  (contact_a_id, contact_b_id, score, status)
           │ ContactMerge   │  (audit trail for merged contacts)
           │ Organization   │  (company records with domain, industry)
           │ TagTaxonomy    │  (hierarchical tag categories)
           └────────────────┘

Key Model Fields

Contact -- full_name, emails[] (PostgreSQL ARRAY with GIN index), phones[], company, organization_id, twitter_handle, twitter_bio, telegram_username, telegram_bio, linkedin_url, relationship_score (0-10), interaction_count, priority_level (high/medium/low/archived), tags[], last_interaction_at, last_followup_at, birthday.

Interaction -- contact_id, platform (email/telegram/twitter), direction (inbound/outbound), content_preview, occurred_at.

FollowUpSuggestion -- contact_id, trigger_type (time_based/event_based/scheduled/birthday/dormant_*), suggested_message, suggested_channel, status (pending/snoozed/dismissed/completed), pool (A/B), snooze_until.

DetectedEvent -- contact_id, event_type (job_change/fundraising/product_launch/promotion/milestone/event_attendance), confidence (0.0-1.0), summary, source_url.

Materialized View

organization_stats_mv aggregates per-organization metrics (contact count, total interactions, average score) and is refreshed hourly via a Celery beat task.

Notable Indexes

ix_contacts_emails_gin -- GIN index on contacts.emails for @> (array containment) queries
ix_contacts_relationship_score -- B-tree for score-based sorting and filtering
ix_contacts_interaction_count -- B-tree for count-based sorting
ix_contacts_full_name -- B-tree for name search

Task Queue

Celery with Redis as both broker and result backend. Tasks use JSON serialization and run in UTC.

Beat Schedule

Task	Schedule	Description
`sync_gmail_all`	Every 6 hours	Sync email threads for all connected Gmail accounts
`sync_google_calendar_all`	Daily 06:00 UTC	Sync calendar events for meeting detection
`sync_telegram_all`	Daily 03:00 UTC	Sync Telegram chat history via MTProto
`poll_twitter_all`	Daily 04:00 UTC	Poll Twitter activity and bio changes
`update_relationship_scores`	Daily 02:00 UTC	Recalculate all contact relationship scores
`generate_suggestions_all`	Daily 08:00 UTC	Run follow-up engine for all users
`refresh_org_stats`	Hourly (:30)	Refresh `organization_stats_mv` materialized view
`reactivate_snoozed_suggestions`	Hourly (:00)	Un-snooze suggestions past their `snooze_until`
`send_weekly_digests`	Monday 09:00 UTC	Email weekly networking digest

Task Safety

Soft time limit: 5 minutes; hard time limit: 10 minutes
Sync failure tasks (notify_sync_failure) create in-app notifications when retries are exhausted
Each task runs its own async event loop via a _run() helper (Celery workers are synchronous)

@shared_task(name="app.services.tasks.sync_gmail_all")
def sync_gmail_all() -> dict:
    return _run(_sync_gmail_all_async())

AI Pipeline

All AI features use the Anthropic Claude API (currently claude-sonnet-4-20250514).

Event Classification

Tweets and bio changes are classified into structured event types:

Tweet/Bio text
    │
    ▼
┌────────────────────┐     ┌─────────────────────────────┐
│ classify_tweet()   │────>│ JSON response:              │
│ classify_bio_      │     │ { event_type, confidence,   │
│   change()         │     │   summary }                 │
└────────────────────┘     └─────────────┬───────────────┘
                                         │
                                         ▼
                              confidence >= 0.7?
                              ├── yes ──> DetectedEvent record
                              └── no ───> discarded

Valid event types: job_change, fundraising, product_launch, promotion, milestone, event_attendance, none.

LLM calls are capped at 5 concurrent requests via asyncio.Semaphore and use exponential backoff with jitter on transient errors (429, 500, 529).

Message Composition

The message composer builds context-aware drafts by assembling:

Contact profile -- name, company, title, bios, relationship score
Conversation history -- last 5 interactions with direction labels
Tone analysis -- analyze_conversation_tone() classifies past messages as formal/casual based on language indicators
Twitter context -- recent tweets fetched via bird CLI (cached 12h in Redis) for time-based and revival triggers
Trigger context -- reason for follow-up (time gap, detected event, birthday, dormant revival)

The prompt instructs the model to produce 2-3 sentences matching the detected tone, referencing the trigger naturally. A template fallback is used if the API call fails.

Auto-Tagging

LLM-generated tags are managed through a TagTaxonomy system with hierarchical categories. Tags can be generated per-contact or applied in bulk via background tasks.

Rate Limiting

Telegram Rate Gate

Telegram's MTProto API enforces aggressive rate limits via FloodWaitError. Ping CRM coordinates across all operations using a Redis-based rate gate:

┌─────────────────┐     FloodWaitError(seconds=N)
│ Any Telegram    │────────────────┐
│ operation       │                ▼
│ (sync, send,    │     ┌──────────────────────┐
│  bio refresh)   │     │ Redis SET            │
└────────┬────────┘     │ tg_flood:{user_id}   │
         │              │ EX = N seconds       │
         │              └──────────────────────┘
         │
         ▼
┌─────────────────┐     ┌──────────────────────┐
│ _check_rate_    │────>│ Redis TTL            │
│ gate(user_id)   │     │ tg_flood:{user_id}   │
└────────┬────────┘     └──────────────────────┘
         │
    TTL > 0?
    ├── yes ──> skip operation / return error with wait time
    └── no ───> proceed

Key behaviors:

Cross-operation coordination -- When any Telegram call triggers FloodWaitError, the gate is set for that user. All subsequent operations (sync, send, bio refresh, group member fetch) check the gate before connecting.
Skip unchanged dialogs -- The sync skips dialogs with no new messages since the last sync timestamp, reducing API calls.
7-day bio freshness filter -- Bio refresh skips contacts whose bio was checked within the last 7 days (telegram_bio_checked_at).
Max contacts per bio sync -- Capped at 100 contacts per run to stay within rate limits.

Frontend Rate Limit UX

The MessageEditor component detects 429 responses and displays a countdown timer. When a send fails with a rate limit, it parses the Retry-After header and shows the remaining wait time to the user.

Identity Resolution

Three tiers of duplicate detection, with a colleague guard to prevent false positives:

Contacts
    │
    ▼
┌────────────────────────────────────┐
│ Tier 1: Deterministic              │
│ - Same email in two contacts       │
│ - Same normalized phone number     │
│ - Email found in Twitter bio       │
│                                    │
│ Result: auto-merge immediately     │
└───────────────┬────────────────────┘
                ▼
┌────────────────────────────────────┐
│ Tier 2: Probabilistic (weighted)   │
│                                    │
│ Signal          Base Weight        │
│ ─────────────── ───────────        │
│ Email domain    0.40               │
│ Name similarity 0.20               │
│ Company match   0.20               │
│ Username sim.   0.10               │
│ Mutual signals  0.10               │
│                                    │
│ Adaptive: weights redistribute     │
│ when signals are unavailable       │
│                                    │
│ Score > 0.85  ──> auto-merge       │
│ 0.70 - 0.85  ──> pending review   │
│ Score < 0.70  ──> ignored          │
└───────────────┬────────────────────┘
                ▼
┌────────────────────────────────────┐
│ Tier 4: Manual Review              │
│ - UI presents side-by-side diff    │
│ - User confirms or rejects match   │
└────────────────────────────────────┘

Blocking Keys

To avoid O(n^2) comparisons, contacts are only compared if they share a blocking key:

Name prefix (first 3 characters)
Name tokens (words >= 3 characters)
Company (exact, lowercase)
Email domain
Email local part tokens (e.g., [email protected] generates tokens pengcheng, chen)
Twitter handle, Telegram username, LinkedIn profile ID

Blocks larger than 500 contacts are skipped to prevent degenerate cases (e.g., the gmail.com domain block).

Colleague Guard

When two contacts share both company and email domain but have clearly different names (similarity < 0.5), the score is capped at 0.35 -- below the 0.70 display threshold. This prevents coworkers at the same company from being flagged as duplicates.

Additional guards:

Short name penalty -- Names shorter than 6 characters (e.g., "Alex", "David") without supporting signals get a 0.5x multiplier
Name-only cap -- When name is the sole available signal, the score is capped at 0.85 to force human review

Merge Behavior

When contacts are merged, the richer contact (by non-null field count) becomes the primary:

List fields (emails, phones, tags) are unioned with order-preserving, case-insensitive dedup
Scalar fields are filled from the secondary contact where the primary has nulls
Interactions are reassigned to the surviving contact
A ContactMerge audit record is created
An IdentityMatch record tracks the merge (secondary's FK becomes NULL after deletion)

Frontend Architecture

Stack

Next.js 15 with App Router (file-based routing under src/app/)
React 19 with server and client components
Tailwind CSS v4 for styling
TanStack React Query for server state management (via QueryClientProvider wrapper)
openapi-fetch for a fully typed API client generated from the backend OpenAPI spec

API Client

The typed client is generated from the backend's OpenAPI schema and uses interceptors for auth:

import createClient from "openapi-fetch";
import type { paths } from "./api-types";

const client = createClient<paths>({ baseUrl: "" });

client.use({
  async onRequest({ request }) {
    const token = localStorage.getItem("access_token");
    if (token) {
      request.headers.set("Authorization", `Bearer ${token}`);
    }
    return request;
  },
  async onResponse({ response }) {
    if (response.status === 401) {
      localStorage.removeItem("access_token");
      window.location.href = "/auth/login";
    }
    if (response.status >= 500) {
      throw new Error(`API error: ${response.status}`);
    }
    return response;
  },
});

Key Pages

Route	Component	Purpose
`/dashboard`	`DashboardPage`	Overview with stats, recent activity, suggestions
`/contacts`	`ContactsPage`	Contact list with search, filters, CSV import
`/contacts/[id]`	`ContactDetailPage`	Full contact profile with interaction timeline
`/suggestions`	`SuggestionsPage`	Follow-up suggestions with inline message composer
`/identity`	`IdentityPage`	Duplicate review queue
`/organizations`	`OrganizationsPage`	Company directory with aggregated stats
`/notifications`	`NotificationsPage`	In-app notification feed
`/settings`	`SettingsPage`	Account connections, sync preferences
`/onboarding`	`OnboardingPage`	Initial setup flow

Error Handling

A class-based error boundary (error.tsx, global-error.tsx) wraps the application. React requires componentDidCatch to be a class method, so this cannot be a function component. The boundary provides recovery options and logs errors for debugging.

Custom Hooks

Hook	Purpose
`useAuth`	JWT token management, login/logout, user state
`useContacts`	Contact CRUD with React Query mutations
`useSuggestions`	Follow-up suggestion fetching and actions
`useDashboard`	Dashboard stats and activity data
`useNotifications`	Notification feed with unread count
`useIdentity`	Identity match review and resolution

System Overview​

Backend Architecture​

API Response Envelope​

Services Layer​

Integrations​

Database Schema​

Key Model Fields​

Materialized View​

Notable Indexes​

Task Queue​

Beat Schedule​

Task Safety​

AI Pipeline​

Event Classification​

Message Composition​

Auto-Tagging​

Rate Limiting​

Telegram Rate Gate​

Frontend Rate Limit UX​

Identity Resolution​

Blocking Keys​

Colleague Guard​

Merge Behavior​

Frontend Architecture​

Stack​

API Client​

Key Pages​

Error Handling​

Custom Hooks​