AI Technology Reference

The AI Behind AMI

AMI is built on over ten distinct AI disciplines — each solving a specific class of problem that ATM teams face. This document explains every technique: what it does, where it’s used, and why it qualifies as AI.

15
AI Techniques
5
Capability Areas
10+
AI Disciplines
4
LLM Providers
💬

Conversational Intelligence

Ask questions in plain English. Get sourced, verified answers in seconds — no SQL, no IT tickets, no waiting.

01

Generative AI — Large Language Models

Deep Learning

Powers the AMI chatbot across five intelligence modes — Dashboard, Analyst, Forensic, Research, and Engineer — each tuned with a different temperature for the task at hand.

  • Google Gemini 3.x (Flash for speed, Pro for depth)
  • Qwen3-Coder-30B on-premise via vLLM
  • NVIDIA Nemotron Super 120B on-premise
  • Anthropic Claude (Haiku for classification, Sonnet for drafting)
  • ATM Advisor: generates strategic reports (ATM Doctor, Morning Brief, Network Advisory)

Large Language Models are deep neural networks trained on billions of tokens of text. They don’t follow hand-coded rules — they learn statistical patterns in language and use those patterns to generate human-quality text, reason over complex contexts, convert natural language to SQL, and produce structured analytical outputs.

The multi-provider architecture means AMI can run entirely on-premise (no data leaves your network) or use cloud models when permitted — and switch between them without changing the user experience.

02

Agentic AI — Autonomous Multi-Step Reasoning

AI Agents

The AMI forensic pipeline is a LangGraph state machine with five specialised nodes that work together autonomously:

  • Planner — decomposes the user’s question into sub-tasks
  • Supervisor — routes each sub-task to the right worker
  • SQL Worker — converts natural language to DuckDB queries
  • Python Worker — runs sandboxed analysis for complex analytics
  • Critic — validates outputs, catches errors, sends work back if needed

A simple chatbot takes a prompt and returns a response. An AI agent decides what to do, not just how to answer. It breaks problems into steps, selects tools, executes a plan, evaluates its own output, and self-corrects.

This is goal-directed autonomous reasoning — the system pursues an objective across multiple steps without human intervention at each stage. It’s the difference between a calculator and an analyst.

03

Natural Language Processing

NLP
  • spaCy pipeline for tokenisation and entity recognition when parsing EJ (Electronic Journal) logs
  • Text mining tool for pattern analysis across transaction log narratives
  • Document intelligence for extracting structured data from unstructured sources
  • Sentiment classification of social media mentions via Claude API

Natural Language Processing is the branch of AI that enables machines to understand, interpret, and generate human language. In AMI, NLP converts messy transaction logs into structured data, classifies social media sentiment, and enables the entire “ask in English” experience.

Without NLP, every query would require SQL expertise. With it, anyone on the team can interrogate the data.

📈

Predictive Analytics

Model how cardholders choose ATMs, segment behaviour patterns, and predict demand shifts — before they happen.

04

Choice Modelling — Multinomial Logit

Statistical ML

The MNL model is calibrated on observed cardholder behaviour to predict which ATM a customer will choose. Seven utility coefficients are learned via Maximum Likelihood Estimation:

  • Driving time (disutility)
  • Site quality score (WCDI)
  • Competition density
  • Capture potential
  • Footfall
  • Branch co-location
  • Site type (wall/drive-thru/lobby)

This feeds a 10-step demand prediction pipeline with bootstrap confidence intervals (500 iterations), maturation curves, and demographic adjustments.

Multinomial Logit is a probabilistic machine learning model. It learns from observed behaviour — the seven coefficients are not hand-coded by analysts but estimated from data using maximum likelihood optimisation.

The model captures the trade-offs cardholders actually make (convenience vs. proximity vs. site quality) and uses those learned trade-offs to predict behaviour at sites that don’t exist yet. This is the core of demand forecasting for new ATM deployments.

05

Clustering & Segmentation

Unsupervised ML
  • DBSCAN — clusters cardholder transactions into activity zones (home, work, transit) using 1 km radius on haversine distances. Up to 6 zones per cardholder.
  • K-Means — creates 6 mobility profiles from H3 hexagon × day-of-week × hour matrices (168 dimensions). Identifies residential, commercial, transit, and mixed-use patterns.
  • Loyalty Segmentation — entropy-based scoring using primary ATM share (50% weight), coefficient of variation (25%), day-of-week entropy (15%), and time entropy (10%). Classifies: Habitual, Transitional, Opportunistic.

Clustering is unsupervised machine learning — the algorithm discovers hidden structure in data without being told what to look for. Nobody tells DBSCAN where home zones are; it finds them from transaction coordinates. Nobody defines the six mobility profiles in advance; K-Means discovers them from the data.

The entropy-based loyalty scoring uses information theory (a foundation of modern AI) to measure how predictable a cardholder’s behaviour is. High entropy = unpredictable = opportunistic. Low entropy = habitual.

06

Regression & Automated Feature Selection

Supervised ML
  • OLS Regression — learns the relationship between POI features (restaurants, banks, malls near an ATM) and transaction volume
  • LassoCV — automatically selects which features matter from 196+ POI candidates. The machine decides, not a human.
  • Ridge Regression — stabilises coefficient estimates when features are correlated
  • Dual-source OLS blend — combines Google Places and Overture Maps POI data into the Weighted Cash Demand Index (WCDI)
  • VIF analysis — detects and removes multicollinear features automatically

Regression is supervised machine learning. The model learns a mathematical function that maps inputs (nearby POIs, population, footfall) to outputs (expected transaction volume) by minimising prediction error on training data.

LassoCV goes further: it automatically performs feature selection by penalising weak predictors to zero. From 196 possible features, the algorithm identifies the handful that actually drive demand. This is automated discovery — the machine finds what matters.

🌍

Site & Spatial Intelligence

Score every potential site, map competitive landscapes, and predict competitor closures — deploy with confidence.

07

Multi-Criteria Scoring with Fuzzy Logic

Decision Science + Optimisation

The Analytic Hierarchy Process (AHP) scores every location on six criteria:

  • WCDI (43.2%) — cash demand from surrounding POIs
  • Footfall (18.3%) — pedestrian traffic (cohort-aware)
  • POI Context (17.6%) — enricher vs. diluter POI groups
  • Competition (11.5%) — nearby competitor ATM density
  • Capture (9.1%) — cardholder recapture opportunity
  • Population (0.2%) — residential demand proxy

Each criterion passes through a logistic S-curve (fuzzy membership function) before weighted aggregation. Validation: Spearman ρ = 0.63 against observed transaction volumes.

Fuzzy logic is a branch of AI that handles uncertainty. Instead of binary “good/bad”, fuzzy membership functions score “how much” a site meets each criterion on a continuous scale. The S-curve shape parameters (steepness k and inflection point) are not guessed — they’re optimised.

Differential evolution jointly optimises all 15 parameters (6 weights + 8 curve shapes + 1 spatial decay) over 500 generations. This is a metaheuristic optimisation algorithm inspired by biological evolution — candidate solutions compete, mutate, and recombine until the fittest survives.

08

Geospatial AI

Spatial Analytics
  • H3 Hexagonal Indexing — Uber’s hierarchical hexagonal grid (Resolution 10, ~1.1 km²) used throughout for spatial aggregation, catchment analysis, and multi-resolution population grids
  • Inverse Distance Weighting (IDW) — competition index: CI = Σ exp(-β × dkm). Nearby competitors weigh more. β is optimised by differential evolution.
  • Gravity Model — fallback demand estimator for greenfield sites with fewer than 30 observations. Demand proportional to attractiveness, inversely proportional to distance.
  • OSRM Routing — driving and walking distance matrices between all ATM pairs within 25 km. Used in MNL utility calculation.
  • Nominatim Geocoding — reverse geocoding for address resolution

Geospatial AI combines spatial data structures, mathematical models, and network algorithms to make location-aware predictions. The system reasons about space the way a human analyst would — “how far is this from competitors?”, “what’s the catchment population?”, “what’s the fastest route?” — but across thousands of locations simultaneously.

The IDW competition index and gravity model are both spatial decay functions with learned parameters. They don’t use hard distance cut-offs — they model the continuous, non-linear relationship between distance and influence.

09

Survival Analysis — Competitor Closure Prediction

Statistical ML

The Cox Proportional Hazards model predicts when a competitor ATM is likely to close, using three learned covariates:

  • Distance to your nearest ATM — closer competitors face more pressure
  • Competitor age — older installations have higher closure risk
  • Area ATM density — oversaturated areas see more closures

Output: closure probability per competitor ATM, classified into high/medium/low risk tiers. Falls back to Logistic Regression when fewer than 30 closure events are available.

Survival analysis is a branch of statistical machine learning that models time-to-event data. The Cox PH model learns which factors accelerate or delay competitor closures from historical data — it doesn’t use business rules or assumptions.

This gives your bank a predictive edge: know which competitor ATMs are likely to close before they do, and position to capture the displaced transactions. The model quality is measured by the c-statistic (concordance) — a standard ML model evaluation metric.

Anomaly Detection & Forensics

Detect transaction shifts, equipment failures, and unusual patterns automatically. Act on signals, not surprises.

10

Time Series Decomposition & Change-Point Detection

Signal Processing
  • STL Decomposition — separates every ATM’s transaction history into trend, seasonal, and residual components using LOESS (locally-weighted regression). Requires 24+ months for full decomposition; falls back to KSA-specific seasonal priors for shorter series.
  • CUSUM (Cumulative Sum Control Chart) — detects structural breaks (level shifts) in the residual signal. Parameters: k=0.5, h=4.0. Backtracks to identify the true break onset.
  • Temporal cross-referencing — detected breaks are correlated with competitor events within ±2 months to establish causation.

STL uses LOESS — a non-parametric machine learning technique that fits local regressions across the data. It doesn’t assume a fixed mathematical form; it learns the shape of the trend and seasonality directly from the data.

CUSUM is a sequential analysis algorithm from the same family used in industrial quality control and financial fraud detection. It detects when a process has shifted — automatically, continuously, across hundreds of ATMs, without a human reviewing each chart.

11

Cascading Anomaly Classification

Anomaly Detection

When an ATM’s monthly transactions deviate significantly (|z| > 2.0), the system classifies the anomaly through a cascading hypothesis test:

  • Downtime — zero transactions, no competitor event nearby
  • Competitive — large negative residual + competitor event within ±2 months
  • Environmental — large observed z-score + high seasonal component
  • Unexplained — significant residual with no structural explanation

Each classification triggers a different operational response.

Anomaly detection is a core AI discipline. But detecting an anomaly is only half the problem — the other half is explaining it. AMI’s cascading classifier doesn’t just flag outliers; it tests hypotheses in sequence, cross-referencing multiple data sources (transactions, competitor events, seasonal patterns) to assign a root cause.

This is automated reasoning over multiple evidence sources — the same approach used in medical diagnosis systems and industrial fault detection.

12

Process Mining & EJ Forensics

Process Intelligence
  • Heuristic Net Discovery (pm4py) — reconstructs actual ATM transaction flows from Electronic Journal event logs
  • Bottleneck Detection — identifies where in the transaction process time is lost
  • Variants Analysis — discovers different process paths and their frequencies
  • Forensic Scanner — detects ghost transactions (cash not taken), cash sensor anomalies, stacked-not-taken events
  • Network Analysis — graph-based pattern detection across transaction sequences

Process mining is AI that discovers actual process flows from event logs. It doesn’t rely on documented procedures — it reconstructs how the ATM actually behaves versus how it should behave.

The heuristic net algorithm automatically identifies the most likely sequence of events, filtering noise and highlighting deviations. This is automated pattern discovery from unstructured operational data — finding the needle in millions of EJ log entries.

🔍

Data Grounding & Verification

Every AI answer is grounded in your actual data. The system searches, verifies, and cites — it does not hallucinate.

13

Retrieval-Augmented Generation (RAG)

Information Retrieval + GenAI
  • Domain documentation (data dictionaries, table schemas, glossaries) is loaded, split by markdown headers, and indexed in a FAISS vector store
  • When a user asks a question, the system retrieves the most relevant documentation passages and injects them into the LLM prompt
  • Domain-specific isolation: ATM, Credit Card, and Debit Card contexts are kept separate to prevent cross-contamination
  • Schema statistics and sample rows are injected for each relevant table

RAG combines two AI disciplines: information retrieval (search) and generative AI (LLMs). Instead of relying solely on what the LLM learned during training, RAG searches your actual documentation and feeds relevant passages to the model.

This grounds every answer in real, current information. The LLM knows your table names, column definitions, and business rules — because it just read them, not because it memorised them months ago.

14

Semantic Embeddings & Golden SQL Library

Representation Learning
  • SentenceTransformer (all-MiniLM-L6-v2) encodes proven SQL query examples into vector representations
  • Cosine similarity finds the most relevant few-shot examples for the current question (minimum threshold: 0.3)
  • The golden SQL library contains verified query patterns for common questions: withdrawal rankings, ghost transactions, downtime spikes, fraud patterns, capacity planning
  • Retrieved examples are injected as few-shot context so the LLM generates SQL that follows proven patterns

Embeddings are neural network representations that capture semantic meaning. “Show me failing ATMs” and “list underperforming machines” map to nearby vectors even though they share no words — because the embedding model learned that they mean the same thing.

This is learned semantic understanding, not keyword matching. The system finds relevant examples based on meaning, not string overlap — which is why it works even when users phrase questions in unexpected ways.

15

Multi-Stage Verification Pipeline

AI Safety

Every answer passes through a four-stage verification pipeline:

  • Plan — the Planner node decomposes the question and identifies required data sources
  • Execute — the SQL/Python workers run the queries and computations
  • Critique — the Critic node validates output: checks for hallucinated columns, broken joins, logical errors, empty results
  • Confirm — only verified output reaches the user; failed checks route back to the Planner for retry

Self-verification is a hallmark of advanced AI systems. The Critic node is an independent AI reviewer — it evaluates the work of other AI components using different criteria than those that generated the output.

This is AI checking AI. The generator optimises for helpfulness; the critic optimises for correctness. The tension between these two objectives is what prevents hallucination and ensures every answer is grounded in real data.

Summary: 15 AI Techniques at a Glance

Every technique, its AI discipline, and the business problem it solves.

# Technique AI Discipline Business Problem Solved
01Large Language ModelsDeep LearningNatural language access to all ATM data
02Agentic AIAI AgentsAutonomous multi-step analysis without human steering
03Natural Language ProcessingNLPParsing logs, classifying sentiment, understanding queries
04Multinomial LogitStatistical MLPredicting which ATM a cardholder will choose
05Clustering & SegmentationUnsupervised MLDiscovering zones, mobility profiles, loyalty tiers
06Regression & Feature SelectionSupervised MLIdentifying which location features drive demand
07Fuzzy Logic + OptimisationDecision ScienceScoring sites on multiple criteria with uncertainty
08Geospatial AISpatial AnalyticsLocation-aware prediction across thousands of sites
09Survival AnalysisStatistical MLPredicting when competitor ATMs will close
10Time Series DecompositionSignal ProcessingSeparating trend from seasonality and noise
11Anomaly ClassificationAnomaly DetectionExplaining why an ATM’s performance changed
12Process MiningProcess IntelligenceDiscovering actual ATM behaviour from EJ logs
13RAGInformation Retrieval + GenAIGrounding AI answers in actual documentation
14Semantic EmbeddingsRepresentation LearningFinding relevant examples by meaning, not keywords
15Multi-Stage VerificationAI SafetyPreventing hallucination; AI checking AI