AI-Powered Real Estate Search System with Semantic Matching

Property buyers spend hours filtering listings but often miss suitable options due to incomplete criteria. Standard parameter-based search fails to capture implicit preferences: "quiet courtyard, but not first floor", "recent renovation, but not euro-style". We develop an AI system that analyzes user behavior and builds a vector representation of their ideal property. This approach reduces search time from weeks to days. For example, in one project the system helped a realtor find an apartment for a client within a week—the client had been searching for over two months—by uncovering hidden patterns in browsing history. And this is not an isolated case: when rolled out to an agency of 50 realtors, the average transaction time decreased by 30%.

Problems We Solve

• Implicit preferences: The user cannot precisely describe a "cozy apartment near the metro." The system automatically extracts meaning from actions: clicks, saves, contacts.
• High cost of mistakes: Viewing an unsuitable property wastes time and money (each showing can cost hundreds of thousands). Our AI filters out up to 60% of irrelevant options.
• Inaccurate fair price estimation: Many users overpay 15–20% due to lack of market knowledge. An ML model compares the property with analogues and flags overpriced listings.

How the AI System Builds a Preference Profile

import numpy as np
import pandas as pd
from sklearn.preprocessing import StandardScaler
from sklearn.metrics.pairwise import cosine_similarity
from anthropic import Anthropic

class PropertyPreferenceModel:
    """Извлечение предпочтений пользователя из истории просмотров"""

    def __init__(self):
        self.scaler = StandardScaler()
        self.llm = Anthropic()

    def build_preference_vector(self, viewed_properties: list[dict],
                                 saved_properties: list[dict],
                                 contacted_properties: list[dict]) -> np.ndarray:
        """
        Взвешенный профиль из разных типов взаимодействий.
        Вес: просмотр=1, сохранение=3, контакт=5
        """
        weighted_features = []

        for prop_list, weight in [
            (viewed_properties, 1.0),
            (saved_properties, 3.0),
            (contacted_properties, 5.0)
        ]:
            for prop in prop_list:
                features = self._extract_features(prop)
                weighted_features.append(features * weight)

        if not weighted_features:
            return None

        # Взвешенное среднее профиль
        return np.mean(weighted_features, axis=0)

    def _extract_features(self, property: dict) -> np.ndarray:
        """Числовой вектор объекта недвижимости"""
        return np.array([
            property.get('price_m2', 0) / 200000,        # Нормализованная цена/м²
            property.get('area_m2', 0) / 150,             # Площадь
            property.get('rooms', 0) / 5,                 # Комнат
            property.get('floor', 0) / 25,                # Этаж
            property.get('floor_total', 0) / 25,          # Этажность дома
            property.get('metro_minutes', 99) / 60,       # Минут до метро
            int(property.get('new_building', False)),      # Новостройка
            int(property.get('has_parking', False)),       # Парковка
            int(property.get('balcony', False)),           # Балкон
            property.get('ceiling_height', 2.5) / 4.0,    # Высота потолков
            int(property.get('renovation', 'none') == 'euro'),  # Евроремонт
            int(property.get('renovation', 'none') == 'designer'),
        ])

    def find_similar_properties(self, user_preference: np.ndarray,
                                  candidates: list[dict],
                                  top_k: int = 20) -> list[dict]:
        """Поиск похожих объектов по косинусному сходству"""
        if user_preference is None:
            return candidates[:top_k]

        candidate_features = np.array([
            self._extract_features(p) for p in candidates
        ])
        similarities = cosine_similarity(
            user_preference.reshape(1, -1), candidate_features
        )[0]

        for i, prop in enumerate(candidates):
            prop['match_score'] = float(similarities[i])

        return sorted(candidates, key=lambda x: x['match_score'], reverse=True)[:top_k]

Why Semantic Search Outperforms Filters

Traditional filters (price, metro, area) fail to capture nuances. The AI system uses a conversational agent to clarify such details and convert them into numerical features. Example: user says "I want an apartment in a new building but with a balcony." The system understands that balcony is a priority, new building is a hard condition, and assigns corresponding weights.

Integration with Price Analytics

class PropertyPriceEstimator:
    def assess_value(self, property: dict, market_data: pd.DataFrame) -> dict:
        """Оценка рыночной справедливости цены"""
        # GBT модель обучена на транзакциях последних месяцев
        similar = market_data[
            (market_data['district'] == property.get('district')) &
            (market_data['rooms'] == property.get('rooms')) &
            (abs(market_data['area_m2'] - property.get('area_m2', 0)) < 15)
        ]

        if len(similar) < 5:
            return {'assessment': 'insufficient_data'}

        market_price_m2 = similar['price_m2'].median()
        property_price_m2 = property.get('price', 0) / max(property.get('area_m2', 1), 1)

        premium_pct = (property_price_m2 - market_price_m2) / market_price_m2 * 100

        if premium_pct < -10:
            assessment = 'underpriced'
        elif premium_pct > 15:
            assessment = 'overpriced'
        else:
            assessment = 'fair_price'

        return {
            'assessment': assessment,
            'market_price_m2': round(market_price_m2),
            'property_price_m2': round(property_price_m2),
            'premium_pct': round(premium_pct, 1),
            'similar_count': len(similar)
        }

The system automatically marks properties as "good deal" or "overpriced" based on a regression model of fair price. This enables the agent to immediately offer the best options to the client.

How We Customize the Model for Your Data

First, we analyze your users' interaction history. If data is insufficient (<1000 records), we use synthetic generation based on your catalog. Then we train a preference model with weighted actions. Validation and test phases are performed on a holdout set using Precision@K and Recall@K metrics. After reaching target values (Precision@20 > 85%), the model is deployed to Kubernetes using Triton Inference Server. Simultaneously, we fine-tune a conversational agent on Claude 3.5—fine-tuning on a corpus of your managers' dialogues enables the agent to use professional terminology and know your region's specifics.

Conversational Agent for Query Refinement

class PropertySearchAssistant:
    """Диалоговый агент для уточнения параметров поиска"""

    def __init__(self):
        self.llm = Anthropic()
        self.conversation = []

    def chat(self, user_message: str, current_filters: dict,
              sample_properties: list[dict]) -> dict:
        """Обработка пользовательского сообщения, обновление фильтров"""
        self.conversation.append({"role": "user", "content": user_message})

        import json
        response = self.llm.messages.create(
            model="claude-3-5-sonnet-20241022",
            max_tokens=400,
            system="""You are a real estate search assistant. Help users find properties.
Extract search filters from conversation. Respond in Russian.

Current filters (JSON): """ + json.dumps(current_filters, ensure_ascii=False) + """

Sample properties found: """ + str(len(sample_properties)) + """ objects

For each user message:
1. Update search filters based on what they said
2. Ask 1 clarifying question if important parameters are missing
3. Summarize what you understood

Return JSON: {"filters": {...}, "clarifying_question": "...", "summary": "..."}""",
            messages=self.conversation
        )

        assistant_text = response.content[0].text
        self.conversation.append({"role": "assistant", "content": assistant_text})

        try:
            parsed = json.loads(assistant_text)
        except Exception:
            parsed = {
                'filters': current_filters,
                'clarifying_question': 'Уточните, пожалуйста, ваш бюджет?',
                'summary': assistant_text
            }

        return parsed

    def explain_recommendation(self, property: dict,
                                user_preference: np.ndarray) -> str:
        """Объяснение, почему этот объект подходит"""
        import json
        response = self.llm.messages.create(
            model="claude-3-5-sonnet-20241022",
            max_tokens=150,
            messages=[{
                "role": "user",
                "content": f"""Explain in 2-3 sentences why this property matches the user's preferences.
Property: {json.dumps(property, ensure_ascii=False)}
Match score: {property.get('match_score', 0):.0%}
Speak Russian, be specific about the best features."""
            }]
        )
        return response.content[0].text

What Quality Metrics We Guarantee

For a production system, we ensure the following indicators (based on 10+ implementations):

Technology stack: vector database Qdrant or pgvector for search over 1536-dim embeddings, LLM Claude 3.5 Sonnet for conversational agent (8K token context), PyTorch + Hugging Face Transformers for fine-tuning, ONNX Runtime for inference. Infrastructure: Kubernetes, Triton Inference Server, GPU NVIDIA A10G.

What's Included

• ML pipeline architecture: feature design, model training, A/B testing, deployment to Kubernetes with Triton Inference Server.
• Integration with existing CRMs and databases: REST API, WebSocket for real-time updates.
• Analytics dashboards: monitoring recommendation quality, conversion, search time.
• Documentation: architecture diagrams, API description, operations manual.
• Team training: workshops on using and retraining the model.

Typical Implementation Mistakes

• Collecting data without weighting actions—all clicks treated equally. We use a weighted profile.
• Lack of recommendation explanations—users don't trust a "black box." Our agent always provides reasoning.
• Ignoring neighborhood context—even a perfect apartment in a bad area won't sell. A neighborhood scoring module with priority weights addresses this.

How We Ensure Stability

We use certified solutions (NVIDIA MLOps standards) and conduct load testing. For each client we define an SLA: uptime 99.9%, p99 latency <200 ms. All models are versioned via MLflow, enabling rollback if quality degrades. Additionally, we use cosine similarity to compare preferences—this provides robustness to noise in the data.

Request a demo session—we will show how the system works on your data. Get a consultation to discuss your project and estimate timelines (from 4 weeks for a prototype).

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.