From Statistics to Predictions in Milliseconds
How Aito combines Bayesian inference with database indexing for instant ML
Aito represents a fundamental shift in machine learning implementation. Instead of training models, it calculates statistics on-demand, enabling real-time predictions without the traditional ML pipeline.
The Bayesian Approach: Why Statistics Beat Models
Traditional ML trains fixed models on historical data. Aito calculates probabilities dynamically using Bayes' theorem, enabling more flexible and explainable predictions.
No Training Phase
Predictions are calculated from raw statistics, not trained models
P(Y|X) calculated directly using indexed conditional probabilities
Automatic Feature Discovery
Bayesian networks discover relationships without manual feature engineering
Mutual information metrics identify predictive relationships
Built-in Uncertainty
Every prediction includes confidence scores for automation decisions
Posterior probabilities provide calibrated uncertainty estimates
The Mathematics Behind Instant Predictions
Aito implements sophisticated Bayesian inference through specialized database indexes
P(Y|X) = P(X|Y) × P(Y) / P(X)How Aito Calculates Statistics Fast Enough for Ad-Hoc Modeling
The key innovation: specialized indexes that make statistical calculations as fast as database queries
Inverted indexes for features
Similar to search engines, but using indexes optimized for statistical operations
Word 'invoice' → items {id1, id2, id3} with countsDatabase optimized for statistical operations
Database contains indexes, caches and precomputation designed for fast inference
Precomputed indexes and offsets over links, row feature vectorsDeep integration between inference and database
Inference uses database primitives directly and maintains low-level caches
Inference operates directly on indexes and maintains specialized segment-level cachesPerformance Optimizations
Three Inference Approaches for Different Use Cases
Aito provides specialized inference methods optimized for different prediction scenarios
Predict: Classification & Regression
Direct Bayesian inference for single-target predictions
{
"from": "invoices",
"where": {
"vendor": "Acme Corp"
},
"predict": "category"
}
Recommend: Collaborative Filtering
Find items that maximize a goal probability
{
"from": "impressions",
"where": {
"user": "alice"
},
"recommend": "product",
"goal": {
"purchased": true
}
}
Generic Queries: Custom Statistics
Flexible statistical queries for complex scenarios
{
"from": "impressions",
"where": {
"user": "alice",
"product.name": {
"$match": "milk"
}
},
"get": "products",
"orderBy": {
"$multiply": [
{
"$similarity": {
"name": "milk"
}
},
{
"$p": {
"$context": {
"click": true
}
}
}
]
}
}
Beyond Traditional ML: Automatic Context Recognition
Aito's representation learning automatically discovers high-level concepts from raw data, eliminating the need for manual feature engineering
Traditional ML requires explicit feature engineering - manually creating features like 'is_weekend' or 'high_value_customer'. Aito automatically discovers these patterns through representation learning based on minimum description length (MDL) principles.
Pattern Recognition
Identifies co-occurring features that form meaningful concepts
Walking like a duck + Swimming like a duck + Quacking like a duck → Concept: "duck"
Automatic Feature Combination
Combines low-level features into high-level representations
{
"$and": [
{ "$has": "rent" },
{ "$has": "real" },
{ "$has": "estate" }
]
}
→ Concept: "real estate rental"
Context-Aware Matching
Recognizes entities and relationships without explicit configuration
"Invoice to Johnson in IT department" → Automatically matches to employee records and department patterns
Benefits of Representation Learning
Intelligent Matching and Prior Knowledge
How Aito uses word matching and Bayesian priors for robust predictions
Prediction target priors
Uses metadata to assign initial probabilities to prediction targets
Technique: Bayesian priors based on prediction target metadata
If customer prefers bread, increase purchase probability for items with bread property
Text match priors
Automatically leverages word and feature matches in inference
Technique: Bayesian priors based on cross-field matching statistics
If description field contains 'Sarah', assume correct processor is Sarah, even if this exact combination hasn't been seen before
Metadata-based priors
Leverages known metadata to inform predictions about new features
Technique: Bayesian priors based on metadata similarity
If user has vegetarian metadata tag, infer vegetarian preferences based on similar users
Smart Use of Prior Knowledge
Bayesian priors prevent overfitting and handle sparse data
New vendor with no history
Approach: Uses category distribution of similar vendors as prior
Benefit: Reasonable predictions even with zero examples
Rare edge case
Approach: Balances specific evidence with general patterns
Benefit: Avoids overfitting to anomalies
Evolving patterns
Approach: Continuous prior updating as data accumulates
Benefit: Adapts to drift without retraining
Advanced Query Operators for Ad-Hoc Feature Engineering
Powerful operators that enable custom feature engineering at query time
$on: Conditional Context
Focus predictions on specific data subsets
{
"from": "invoices",
"where": {
"region": "EU",
"$on": [
{
"amount": {
"$gt": 10000
}
},
{ "region": "EU" }
]
},
"predict": "approval_time"
}
Use case: Segment-specific predictions without separate models
$knn: K-Nearest Neighbors
Find similar items using weighted feature similarity
{
"from": "products",
"where": {
"$knn": {
"k": 10,
"near": {
"name": "Acme CRM",
"category": "software"
}
}
}
}
Use case: Similarity-based recommendations and matching
$numeric: Smart Binning
Automatic discretization of continuous variables
{
"from": "customers",
"where": {
"age": { "$numeric": 25 }
},
"predict": "churn_risk"
}
Use case: Handle numeric features without manual binning
Combining Operators for Complex Logic
{
"from": "transactions",
"where": {
"$on": [
{
"$knn": {
"k": 5,
"near": { "description": "AWS Consulting" }
},
"amount": { "$numeric": { "$gte": 1000 } }
}
},
{ "region": "EU" }
]
},
"predict": "fraud_risk"
}
Build sophisticated feature engineering logic without code
Performance Validation: Benchmarks and Evaluations
Rigorous testing demonstrates Aito's performance compared to traditional ML approaches
Accuracy Comparison Across Datasets
Comprehensive benchmarks across multiple UCI datasets compared to traditional ML methods
| Dataset | Aito | Random Forest | SVM | Insight |
|---|---|---|---|---|
| Spam Detection | 95% | 97% | 87% | Competitive with best ensemble method |
| Splice Junction | 93% | 96% | 94% | Strong performance on genomic data |
| Shuttle | 98% | 99% | 78% | Near-optimal multi-class classification |
| Invoice Processing | 98% | 99% | 26% | Excellent real-world business performance |
Aito achieves 96% mean accuracy across benchmarks, ranking 3rd out of 8 methods while requiring zero feature engineering
Enterprise-Scale Performance Metrics
Query Response Time
20-200ms
Even at 10M+ record scale
Massive Scale Processing
Sub-200ms at 10M records
Linear scalability with 100% reliability
High-Accuracy Predictions
84-91% at scale
Better accuracy with larger datasets
Memory Efficiency
150-300MB for 10M records
Predictable resource scaling
Feature Engineering
Automatic
vs weeks of manual work
Model Management
Zero maintenance
vs continuous retraining pipelines
Production Deployments
Posti Group
Purchase invoice automation
Helsingö
Product catalogue matching (IKEA to Helsingö)
Jokiväri
Purchase invoice processing
Try It Yourself: Interactive Query Examples
Explore different query types with our interactive widget
Try It Yourself: Interactive Query Examples
Explore different query types with our interactive widget
Prediction with Confidence
Classify with calibrated confidence scores
{
"from": "invoices",
"where": {
"Description": "Aws Cloud"
},
"predict": "Processor",
"select": [
"$p",
"Name",
"Department",
"Role"
]
}Address
Episto Oy
Putouskuja 6 a 2
01600 Vantaa
Finland
VAT ID FI34337429
