SPAM-JAMR

Threat Intelligence Engine
A multi-layered, on-device machine learning system that analyzes 47 behavioral features across 8 data sources to protect you from spam, robocalls, and phone scams — all without ever sending your data off your phone.
ADAPTEDGEML ENGINE v1.0
THREE-TIER ARCHITECTURE
How the engine is structured

SPAM-JAMR's threat engine is built as a three-tier pipeline. Each tier has a single responsibility, and data flows forward from raw sources to a final risk verdict. No tier reaches back — information always moves in one direction.

TIER 1 — COLLECT

DataiNetX™

The Data Intelligence Network. Gathers raw intelligence from 8 distinct sources, detects attack scenarios (neighbor spoofing, rotating numbers, robocall farms), and transforms everything into a normalized 47-element FloatArray for the ML engine.

TIER 2 — PREDICT

AdaptEdgeML™

The Machine Learning Engine. Receives the 47-feature array, runs it through a single-layer neural network with trained weights and sigmoid activation, and outputs a risk probability. Learns from every user action — gets smarter with use.

TIER 3 — ANALYZE

ThreatIntelligence

The Decision Layer. Takes the ML prediction and combines it with three other risk factors in a weighted analysis. Then checks for escalation triggers that can bump the risk level up. Produces the final score and verdict.

📡
INTERACTIVE PIPELINE
Tap a scenario to see how the engine responds
SELECT A CALL SCENARIO
📞
SUSPECTED SPAM
Unverified caller, out-of-area, FTC complaints on file
MEDIUM → HIGH
LEGITIMATE CALL
Known contact, verified by carrier, local area code
MINIMAL
UNKNOWN CALLER
First contact, verified, no complaints, local number
LOW
🤖
ROBOCALL ATTACK
Rotating numbers, rapid calls, sequential pattern, not verified
CRITICAL
🎭
NEIGHBOR SPOOF
Matching area code and prefix, but not in contacts
HIGH
An unverified caller from an out-of-area code with FTC complaints. The engine detects NOT_VERIFIED and escalates from MEDIUM to HIGH.
RAW DATA SOURCES FTC Reports FCC Reports SpamDatabase Area Code DB STIR/SHAKEN Call History Community Contextual Intel DataiNetX™ CENTRALIZED DATA AGGREGATION — aggregateData() ftcComplaints fccReports databaseMatches areaCodeRisk patternScore communityReports contextual • scenarios ATTACK ↓ AggregatedData (RawDataSources + FloatArray[47] + detectedScenarios) ThreatIntelligence THE ORCHESTRATOR — analyzeThreat() CSS → configureEngines(mlEngine, dataiNetX) • routes AggregatedData to all factors F1: AREA CODE 25% WEIGHT rawData.areaCodeRisk FTC metro complaint density Local vs out-of-area scoring Known scam corridors 55% ↓ RiskFactor F2: ML PREDICTION 35% WEIGHT FeatureExtractors 47 Features → 5 Categories → FloatArray AdaptEdgeML™ forwardPass(f × w) • sigmoid → RiskPrediction 65% ↓ RiskFactor F3: BEHAVIOR 20% WEIGHT rawData.previousBlocks rawData.communityReports rawData.isWhitelisted rawData.shortCallCount ratio Rapid calls • Rotating nums 50% ↓ RiskFactor F4: REGULATORY 20% WEIGHT rawData: FTC 30% • FCC 30% ftcComplaints • fccReports STIR/SHAKEN 40% PASS • FAIL • NOT_VERIFIED 45% ↓ RiskFactor × 0.25 × 0.35 × 0.20 × 0.20 WEIGHTED SUM → BASE SCORE 58 ESCALATION ENGINE collectEscalationTriggers() NOT_VERIFIED RAPID_CALLS ROTATING_NUM OFF_HOURS SEQ_PATTERN PREV_BLOCKED HI_RISK_AREA +1 LEVEL RISK MINLOWMEDHIGHCRIT FINAL VERDICT 73 HIGH MEDIUM → HIGH (ESCALATED) NOT_VERIFIED +1 LVL RAPID_CALLS +0 (CAP) ↓ TO BDE AdaptEdgeCal PRE-TRAIN • 500K samples on first launch BlockingDecisionEngine TIERED DECISION WATERFALL T0 EMERG T1 TRUST T1.5-2 PREF T3 STIR T4 ML/TI T4.5 RISK T5 UNKNOWN BLOCK / ALLOW 911 PASSWL/ContactUser TogglesVerifyScoreThresholdShow CTA SIFI UserSettingsPresetsThresholds USER ACTION → autoLearn() → RETRAIN WEIGHTS 🔔 Notification → User taps action → triggers autoLearn() NotificationActionReceiver + ReportSpamDialog ● BLOCKED ×1.5 ● REPORTED ×2.0 ● WHITELISTED ×2.0 ● CALLED_BACK ×1.5 Retrains AdaptEdgeML™ FEATURE_WEIGHTS[47] SOURCES → DataiNetX™ AGGREGATION → TI 4-FACTOR ANALYSIS → ESCALATE → BDE TIERS → ML LEARNS SPAM-JAMR™ ENGINE ARCHITECTURE • ALL ON-DEVICE • ZERO EXTERNAL API CALLS

SCROLL HORIZONTALLY TO VIEW FULL DIAGRAM • PINCH TO ZOOM

🧬
47 ML FEATURES
What the engine sees in every call

Every incoming call is decomposed into 47 distinct numerical features organized across 5 categories. DataiNetX™ extracts these from the raw data sources and normalizes them into a FloatArray — the language AdaptEdgeML™ understands. Each feature is a signal. Alone, any single feature means little. Combined through weighted multiplication, patterns emerge that no single rule could catch.

15
Temporal Patterns
Call duration, frequency, burst patterns, time-of-day analysis, weekend ratio, peak spam hours, night calls, time between calls, daily maximums
10
Number Analysis
Area code matching, prefix similarity, toll-free/premium detection, VOIP indicator, sequential digits, number length anomaly, format validation
8
Contextual Intel
First contact detection, recency scoring, geographic distance, Wangiri ring analysis, callback pressure, number entropy, timezone mismatch, neighbor spoofing
8
User Behavior
Answer ratio, block history, spam reports, contact list match, callback attempts, voicemail presence, conversation duration, repeat call patterns
6
Community & Database
Local spam reports, FTC complaint count, FCC complaint count, global blocklist match, area code risk score, database match confidence
ATTACK SCENARIO DETECTION

Before features reach the ML engine, DataiNetX™ runs scenario-level pattern matching to detect coordinated attack patterns that individual features might miss:

4-FACTOR RISK ANALYSIS
How the final score is calculated

ThreatIntelligence doesn't rely on any single signal. It combines four independent risk assessments, each weighted by reliability and predictive value, into a single composite score from 0 to 100.

AREA CODE RISK
25%
Geographic intelligence from NANPA area code data. Cross-references caller area codes against FTC/FCC complaint density, known scam corridors, and local vs. out-of-area status.
ML PREDICTION
35%
AdaptEdgeML™'s neural network output. The heaviest weight because ML adapts to new spam tactics that static rules miss. Sigmoid activation normalizes output to 0-1 probability.
BEHAVIOR ANALYSIS
20%
Real-time behavioral signals: call frequency, timing anomalies, burst calling detection, and historical interaction patterns between this number and the user.
REGULATORY DATA
20%
Combines: STIR/SHAKEN verification status (40% of this factor), FTC complaint data (30%), and FCC complaint data (30%).
RISK LEVEL SCALE
MINIMAL 0-24
LOW 25-44
MED 45-64
HIGH 65-79
CRIT 80-100

The weighted sum maps to one of five risk levels, each with a recommended action.

🚨
ESCALATION ENGINE
When the base score isn't enough

After ThreatIntelligence calculates the base risk score, the Escalation Engine checks for specific behavioral triggers that indicate higher danger than the raw numbers suggest.

7 ESCALATION TRIGGERS
TriggerDetectionThreshold
NOT_VERIFIEDSTIR/SHAKEN failed or unverifiedAlways +1 level
RAPID_CALLSSame number calling repeatedly2+ in 5 min or 5+ in 30 min
ROTATING_NUMSimilar numbers with sequential digits2+ matching prefix
SEQ_PATTERNRobocall farm detection3+ distinct sequential numbers
PREV_BLOCKEDNumber was previously blockedIn blocked list
OFF_HOURSCalling outside normal hoursBefore 8 AM or after 9 PM
HI_RISK_AREAKnown high-spam area codeArea code risk score 70+
ESCALATION RULES
NOT_VERIFIED present+1 level (guaranteed)
3+ triggers (incl. NOT_VERIFIED)+2 levels
2+ triggers (no NOT_VERIFIED)+1 level
FAILED verificationAlways HIGH min
MAXIMUM ESCALATION: 2 LEVELS
🎯
ADAPTIVE LEARNING
The engine learns from you

AdaptEdgeML™ doesn't just follow static rules — it learns from every action you take. When you block a number, report spam, or trust a caller, the engine adjusts its internal weights through gradient descent.

HOW IT LEARNS

The neural network uses a single-layer perceptron with sigmoid activation. Training uses stochastic gradient descent with a learning rate of 0.01.

Your ActionSignalWeightWhy
Block a numberSPAM1.5xStrong spam signal
Report spamSPAM2.0xStrongest — explicit confirmation
Quick hang-upSPAM0.8xWeak — could be accidental
Ignore repeatedSPAM0.7xWeak — passive rejection
WhitelistSAFE2.0xStrongest safe — explicit trust
Answer (full call)SAFE1.2xModerate — user engaged
Call backSAFE1.5xStrong safe — voluntary contact
TRAINING LIFECYCLE

Weights persist in local storage and survive app restarts.

🔒
PRIVACY BY DESIGN
Your data never leaves your phone
🏠
100% On-Device Processing

All 47 features extracted and all ML predictions run entirely on your device. No cloud servers, no API calls for threat data.

🧠
Local Model Training

Neural network weights trained and stored exclusively in your phone's local storage.

📵
No PII in Logs

Release builds strip all personally identifiable information. Phone numbers and contact names never in system logs.

📡
Offline-Capable

Threat databases downloaded and stored locally. Full call screening works offline after initial setup.

🔐
Carrier Verification Only

STIR/SHAKEN status comes from your carrier's network — SPAM-JAMR reads it, never sends it.