Smart Automated Notifications: Trigger System with AI Optimization

Problem: Lost Sales from Untimely Communications

An e-commerce store with 50,000 daily visitors loses up to 70% of potential sales due to delayed communications. Dozens of abandoned carts every day, hundreds of users leave without buying. Manual mailings can't keep up—the moment is lost. We develop an AI trigger communication system that analyzes each user's behavior in real-time and sends a personalized message via the right channel at the optimal time. Result: conversion from triggers increases 2-4x, retention improves by 30%.

The ML-based trigger system solves three tasks: channel selection (email/SMS/push), send time determination, and content generation. Unlike static rules, AI adapts to each user—their activity patterns, preferences, and purchase history. We use LLMs (Claude, GPT-4) to generate text that sounds like a real manager, not a template mailing. The system achieves open rates up to 55% for abandoned carts and CTR of 15-25% for personalized push. AI is 3-5 times more effective than rule-based approaches for these metrics.

How the ML Model Selects Channel and Send Time?

The ML model considers many factors: historical channel effectiveness for the event type, user preferences, current channel load. For example, email works best for abandoned carts (open rate 40-55%), while push or SMS is better for urgent price change notifications. Send time is determined based on activity patterns: if a user opens emails in the evening, the system sends around 6 PM. The model trains on historical interaction data using features: hour of day, day of week, event type, previous responses. We use gradient boosting (CatBoost) for send time regression and channel classification.

Code: Base System Class with Channel and Time Selection

from anthropic import Anthropic
import pandas as pd
import numpy as np
from dataclasses import dataclass
from enum import Enum
import json

class Channel(Enum):
    EMAIL = "email"
    SMS = "sms"
    PUSH = "push"
    IN_APP = "in_app"

@dataclass
class TriggerEvent:
    user_id: str
    event_type: str
    event_data: dict
    timestamp: float

class TriggerCommunicationSystem:
    def __init__(self):
        self.llm = Anthropic()
        self.send_time_model = None
        self.channel_model = None

    def process_trigger(self, event: TriggerEvent, user_profile: dict) -> dict:
        """Full trigger processing: channel + time + content"""
        channel = self._select_channel(user_profile, event.event_type)
        send_delay_hours = self._optimal_send_time(user_profile, event.event_type)
        content = self._generate_content(event, user_profile, channel)
        if self._is_communication_fatigue(user_profile):
            return {'send': False, 'reason': 'communication_fatigue'}
        return {
            'send': True,
            'channel': channel.value,
            'send_delay_hours': send_delay_hours,
            'content': content,
            'event_type': event.event_type
        }

    def _select_channel(self, user: dict, event_type: str) -> Channel:
        preferred = user.get('preferred_channel')
        if preferred:
            return Channel[preferred.upper()]
        channel_effectiveness = {
            'abandoned_cart': {'email': 0.15, 'push': 0.08, 'sms': 0.12},
            'inactivity': {'email': 0.05, 'push': 0.06, 'sms': 0.04},
            'price_drop': {'push': 0.12, 'email': 0.10, 'sms': 0.08},
            'order_shipped': {'sms': 0.25, 'email': 0.20, 'push': 0.18},
        }
        effectiveness = channel_effectiveness.get(event_type, {'email': 0.1})
        best_channel = max(effectiveness, key=effectiveness.get)
        return Channel[best_channel.upper()]

    def _optimal_send_time(self, user: dict, event_type: str) -> float:
        active_hours = user.get('active_hours', list(range(9, 22)))
        if event_type == 'abandoned_cart':
            return 1.5
        elif event_type == 'price_drop':
            return 0.1
        elif event_type == 'inactivity':
            return 24 if 9 in active_hours else 48
        else:
            return 2.0

    def _generate_content(self, event: TriggerEvent, user: dict, channel: Channel) -> dict:
        channel_constraints = {
            Channel.SMS: {'max_chars': 160, 'format': 'plain'},
            Channel.PUSH: {'max_chars': 100, 'format': 'title+body'},
            Channel.EMAIL: {'max_chars': 2000, 'format': 'html'},
            Channel.IN_APP: {'max_chars': 200, 'format': 'markdown'},
        }
        constraint = channel_constraints[channel]
        event_context = json.dumps(event.event_data, ensure_ascii=False)[:300]
        response = self.llm.messages.create(
            model="claude-3-5-sonnet-20241022",
            max_tokens=300,
            messages=[{
                "role": "user",
                "content": f"""Generate a {channel.value} message for trigger event.

Event: {event.event_type}
Event data: {event_context}
User name: {user.get('first_name', 'Customer')}
User purchase history: {user.get('total_orders', 0)} orders, avg order ${user.get('avg_order_value', 0):.0f}
Channel: {channel.value} (max {constraint['max_chars']} chars)

Return JSON:
{{
  "subject": "email subject or push title",
  "body": "message body",
  "cta": "call to action text",
  "cta_url": "URL path"
}}

Tone: friendly, personal. Mention specific item if available. No generic marketing language."""
            }]
        )
        try:
            return json.loads(response.content[0].text)
        except Exception:
            return {
                'subject': f"We have something for you, {user.get('first_name', '')}!",
                'body': response.content[0].text[:constraint['max_chars']],
                'cta': 'View Now'
            }

    def _is_communication_fatigue(self, user: dict) -> bool:
        messages_last_7d = user.get('messages_received_7d', 0)
        opens_last_7d = user.get('messages_opened_7d', 0)
        if messages_last_7d >= 5:
            return True
        if messages_last_7d >= 3 and opens_last_7d == 0:
            return True
        return False

A/B Testing Trigger Messages: Compare Variants

To find optimal strategies, we run A/B tests across channels, timing, and content. Results update in real time—you can adjust parameters on the fly. For example, a test may show that push with emojis yields 20% higher CTR than without.

class TriggerABTest:
    """Testing variants of trigger messages"""

    def __init__(self, test_name: str, variants: list[dict]):
        self.test_name = test_name
        self.variants = variants
        self.results = {v['name']: {'sent': 0, 'opened': 0, 'clicked': 0, 'converted': 0}
                        for v in variants}

    def assign_variant(self, user_id: str) -> dict:
        """Deterministic variant assignment"""
        idx = hash(f"{self.test_name}_{user_id}") % len(self.variants)
        return self.variants[idx]

    def compute_results(self) -> dict:
        results_summary = {}
        for variant_name, stats in self.results.items():
            sent = stats['sent']
            if sent == 0:
                continue
            results_summary[variant_name] = {
                'open_rate': stats['opened'] / sent,
                'click_rate': stats['clicked'] / sent,
                'conversion_rate': stats['converted'] / sent,
                'sample_size': sent
            }
        return results_summary

Channel Effectiveness Comparison

Rule-based vs ML: Which is More Effective?

Rule-based works for simple scenarios with low variability. ML is for complex, personalization-critical cases. In practice, we often combine them: rule-based as a fallback, ML as the main engine.

Key Advantages of the AI System

LLM-driven content personalization—the model generates text considering purchase history, current browsing, and user name. Open rates are 2-3x higher than bulk mailings. ML time optimization—the model analyzes when a specific user is active and sends the message at the moment of maximum read probability. Frequency control—a built-in communication fatigue detector prevents sending more than 3-4 messages per week, reducing unsubscribes. As a result, retention increases and spam complaints drop 5-10x. Using RAG message generation, we fetch relevant product data for each user, making content even more personalized. MLOps practices ensure the system is reliable and scalable.

What's Included in System Development?

1. Analytics — audit of current communications, collection of user activity patterns.
2. Design — event-driven system architecture, stack selection (PyTorch, Hugging Face, ChromaDB).
3. Development — implementation of ML modules for channel and time selection, content generation via LLM.
4. Integration — connection to CRM (Bitrix24, AmoCRM), ESP (SendGrid, Unisender) and SMS gateways.
5. A/B Testing — experiments to find optimal strategies.
6. Documentation & Training — API docs, instructions for marketers.

We have 5+ years of experience in ML systems for e-commerce, having delivered over 30 projects. We use a modern stack: PyTorch, LangChain, Triton Inference Server. We guarantee measurable results—we track open rate, conversion, and compare against baseline. Implementation typically pays back within 2-3 months, generating $50,000 to $200,000 in additional annual revenue for an average e-commerce project.

We'll assess your project in 1-2 days—contact us for a consultation. Get a cost and timeline estimate tailored to your needs. The additional revenue from the AI system can exceed costs by 5-10x in the first year.

Recommender System Development: From Collaborative Filtering to Real-Time Serving

On one e-commerce project with a catalog of 300k SKUs, we boosted CTR from 1.8% to 4.4% — a 2.4x increase. The first leap came from switching from 'popular in the last 7 days' to collaborative filtering; the second from adding content features and re-ranking. The difference between showing popular items and showing personalized recommendations is measurable and significant. Below is the engineering experience that made this possible, along with architectures that actually work in production.

Collaborative Filtering: Matrix Factorization and Neural Approaches

Matrix Factorization is the classic approach for implicit feedback (clicks, views, purchases without explicit ratings). ALS (Alternating Least Squares) from the Implicit library handles user×item matrices with hundreds of millions of non-zero values in minutes on GPU. Latent factors 64–256, regularization λ=0.01–0.1 are starting parameters. Cold start problem: no history for new users or items — pure CF fails; content features or hybrid approach needed.

Neural Collaborative Filtering (NCF) replaces the dot product with a neural network. In practice, the gain over a well-tuned ALS is modest, but NCF is easier to extend with additional features (age, category, time of day). Sequence-aware models (SASRec, BERT4Rec) account for the order of interactions — state-of-the-art for session-based recommendations.

How to Choose Recommender System Architecture?

The answer depends on data, load, and cold start requirements. Below are three main approaches with selection criteria.

Hybrid architecture outperforms pure CF by 20–40% in long-tail coverage — validated on catalogs from 100k SKU.

Content-Based Filtering: When Interaction History is Scarce

Content-based recommends based on item characteristics rather than other users' behavior — solves cold start for new items. Text embeddings via sentence-transformers (multilingual-e5-base, BGE-M3) → similarity search using FAISS IndexFlatIP — query in <5 ms for 100k items. Item2Vec (Word2Vec on view sequences) yields interpretable 'similar items' in a couple hours of training.

Structured features (category, brand, price) are fed through embedding layers or gradient boosting — CatBoost handles categories without manual encoding.

Why Hybrid Models Work Better?

Production systems are almost always two-level. Stage 1 (Retrieval) — fast selection of 100–500 candidates from 300k items using ALS or Two-Tower model with vector search (FAISS, Qdrant). Stage 2 (Ranking) — heavy ranker on LightGBM or neural network with cross-features, time, device, and session context. LightFM is a good starting point for medium scale without heavy infrastructure. Our practice shows: moving from single-stage to two-stage yields a 15–25% accuracy improvement with only 20–30 ms additional latency.

Real-Time Serving: Architecture Under Load

Latency SLA — 50–100 ms at thousands of requests per second. Base recommendations precomputed (batch job hourly) → Redis by user_id → <5 ms. Real-time re-ranking via Kafka for events (clicks, cart adds) → update of context features. Feature serving — Redis with TTL (views in 24 hours, last clicked item). At 10k req/s, we deploy Redis Cluster with replication.

A/B testing is the only reliable way to measure improvements. Offline metrics do not always correlate with online. Kohavi et al., 'Online Controlled Experiments at Large Scale' (KDD 2013) — a must-read for the team. Test on 5–10% of traffic, monitor CTR, conversion, revenue per session. One of our client systems after hybridization increased revenue by 18% over a month of A/B.

Recommender System Development Timeline

The stages and typical time frames are in the table below. Costs are calculated individually based on catalog scale and latency requirements.

What's Included in Turnkey Development

1. Data audit — user×item matrix density (typically <0.1%), activity distribution, temporal patterns, cold start statistics.
2. Baseline — 'popular' as a simple threshold that is often hard to beat.
3. Iterative improvement — ALS → content features → two-stage → sequence-aware. Each step with A/B.
4. Serving infrastructure — batch precomputation, Redis, real-time re-ranking, Grafana monitoring.
5. Documentation — model card with metrics, deployment instructions, feature descriptions.
6. Team training — session on interpreting results and model fine-tuning.
7. Support — 1 month post-launch (incident fixes, pipeline tuning).

We are a team with 7+ years of experience in recommender systems, having delivered over 30 projects for e-commerce and media. We guarantee transparent A/B testing and documented metric improvements.

Want to assess the growth potential of your catalog? Contact us for a free data audit. Order recommender system development — first prototype within two weeks.

from implicit.als import AlternatingLeastSquares

model = AlternatingLeastSquares(
    factors=64,
    regularization=0.05,
    iterations=15,
    use_gpu=True
)
model.fit(user_item_matrix)

More about the mathematics of recommender systems — in specialized literature.