Click any layer to explore its interactive tools
DATA SPINE Unified Intelligence 3 platforms → anomaly correlation DIAGNOSTIC BRAIN Quality Intelligence 5-gate classification engine PREDICTIVE SHIELD Predictive Intelligence Forecasting + drift detection signals → classifications → forecasts →

Compounding Returns

Better classification → better coaching → fewer failures → more resources for systemic prevention → reduced volume → improved SL without headcount. The system gets better the longer it runs.

Layer Interdependence

The diagnostic brain needs the data spine's cross-platform visibility. The predictive layer needs historical baselines. No vendor spans all three boundaries.

Decision Cycle Compression

Traditional: Monday review → Tuesday analysis → Wednesday coaching → Friday hope. This architecture: same-day response by eliminating sequential dependencies.

Zero Capex Stack
Python 3.11, MCP SDK, asyncio, pandas, scipy, Redis, PostgreSQL, n8n. Runs against existing API endpoints already in platform licensing.
Domain Agnostic
Swap CCaaS, WFM, or ticketing vendors — write one connector, zero consumer changes. N+M complexity instead of N×M.
ML Bootstrap Path
Gate-based rules generate labeled training data. After 12+ months: supervised model trains on labels no one else in the industry has.
Connector Map
MCP Architecture
Anomaly Simulator
Financial Impact
Click a platform to inspect API endpoints — or click "Add Platform" to see integration scaling
GC
Genesys Cloud
CCaaS — Call routing, queue metrics, QA evaluations
Endpoints
4
Fastest Poll
5 min
Auth
OAuth2
Rate Limit
300/min
UK
UKG
WFM — Schedules, adherence, timecards, labor costs
Endpoints
3
Fastest Poll
daily
Auth
API Key
Rate Limit
60/min
HD
Helpdesk
Tickets — Escalations, categories, resolution tracking
Endpoints
2
Fastest Poll
real-time
Auth
Bearer
Rate Limit
120/min
+
Add Platform
Any CCaaS, WFM, or Ticketing System

Traditional Integration (N×M)

3 platforms × 18 consumers → 4 × 18
54
+18 new integration points
vs

MCP Integration (N+M)

3 connectors + 18 tools → 4 + 18
21
+1 new connector
Adding one platform requires 1 new connector. Zero consumer changes. Integration timeline: weeks, not months.
Data Flow — Sources → Intelligence Layer → Outputs
Sources
3 Platforms
9 API endpoints
extract → normalize → correlate
MCP Server
Intelligence Layer
Agent ID resolution <2ms
Outputs
18 Reports
DPR, scorecard, cost, QA
Why MCP Abstraction
Direct API wrappers create N×M complexity. MCP lets consumers describe what they need — the server handles extraction.
Agent ID Resolution
Genesys: UUID. UKG: employee number. Helpdesk: email. Redis cache maps any ID to canonical in <2ms.
Platform-Agnostic
Swap Genesys for NICE CXone, add Cognigy and Observe.AI — the MCP layer remains identical.
MCP Server — Tool Registration Pattern
mcp_server.py — daily performance tool
from mcp.server import Server
from mcp.types import Tool, TextContent

server = Server("ops-intelligence")

@server.tool()
async def get_daily_performance(date: str) -> list[TextContent]:
    """Fetch and correlate daily metrics across all platforms."""
    ccaas, wfm, tickets = await asyncio.gather(
        genesys.get_interactions(date),
        ukg.get_timecards(date),
        helpdesk.get_tickets(date)
    )
    perf = build_unified_model(ccaas, wfm, tickets)
    anomalies = detect_anomalies(perf, baseline=get_rolling_30d(date))
    return [TextContent(text=perf.to_report(anomalies))]
Technology Stack
Python 3.11MCP SDKasyncioGenesys Cloud APIUKG APIREST / WebhooksRedisPostgreSQLpandasscipy (Erlang C)n8n
Live Anomaly Correlation — Adjust Metrics to See Classification
AHT (deviation)+0%
Ticket Volume (vs. baseline)1.0x
Schedule Adherence92%
Call Volume (deviation)+0%
Service Level80%
FCR (deviation)0%
NORMAL OPERATIONS
All metrics within baseline parameters
No anomaly detected. Cross-platform correlation shows aligned signals.
AHTNORMAL
TicketsNORMAL
AdherenceNORMAL
VolumeNORMAL
Service LevelON TARGET
FCRNORMAL
Why Multi-Signal
Single-metric alerts = noise. AHT↑ + Tickets↑ + Adherence→ is a system latency signature.
Erlang C Integration
CAPACITY anomaly triggers automatic staffing recalculation using live data.
Action Routing
SYSTEMIC → platform team. STAFFING → WFM. CAPACITY → Erlang recalc. Classification determines who gets the alert.
Input your center's parameters — see quantified savings in real time
Total Agents120
Blended Hourly Rate$28
Current Utilization54%
Target Utilization60%
Current AHT (min)10.7
Target AHT (min)10.2
Annual Turnover45%
Replacement Cost / Agent$7,000
Utilization Savings
$360K
6 pts × 1.5 FTE equiv
AHT Savings
$93K
30s = 4.7% efficiency
Turnover Cost
$378K
54 departures × $7K
Turnover Reduction (5%)
$42K
6 fewer replacements/yr
Combined Annual Opportunity
$495K
7.1% of annual labor spend
Board-Validated
Each utilization point ≈ 1.5 FTE ≈ $40K. CFO independently calculated — matched this model exactly.
AHT Economics
Each 30-second AHT improvement = $93K annual savings at default parameters.
Data Spine Required
Cost per contact needs both labor cost (WFM) and volume (CCaaS). Without the unified data layer, this is manual and monthly.
12
API Endpoints
<2ms
ID Resolution
18
Auto Reports
104wk
Trend Depth
20+hrs
Saved / Week
Scenario Simulator
Classification Engine
Agent Scoring
Architecture Decisions
Select a QA failure scenario — watch the engine process through all five gates
15 agents fail "accurate information"
Standard QA: schedule 15 coaching sessions. But should you?
Failing: 15Criterion: Q7
AHT spikes 40% on Tuesday
Supervisors flag agents as "taking too long." 30+ coaching plans queued.
AHT: +40%Affected: all
Agent scores 62% on empathy
Single agent, consistent low score. Peer average is 88%.
Failing: 1Peer avg: 88%
5 agents fail documentation — same hire month
All hired Oct 2024. Same trainer. Individual plans generated.
Failing: 5Cohort: Oct 24
12 agents fail hold-time compliance
Exceeding max hold during interactions. Compliance team flagging.
Failing: 12Avg hold: 4.2m
2 agents miss call opening script
Flagged across multiple evaluations. 96% of peers pass.
Failing: 2Peer rate: 96%
Build Your Own Scenario
Define custom parameters and run them through the five-gate engine.
Agents failing
Total evaluated
Gate 1 — System health
Platform outage / latency spike
Gate 3 — Process currency
KB / SOP updated within 7 days
Gate 4 — Routing accuracy
Skill mismatch detected
Gate 5 — Cohort
Shared attribute
GATE 01
System Health Check
Platform outages, latency during eval window?
GATE 02
Pattern Width
3+ agents fail same criterion? Rate > 25%?
GATE 03
Process Currency
SOP or KB updated within 7 days?
GATE 04
Routing Accuracy
Interactions correctly skill-matched?
GATE 05
Cohort Analysis
Shared hire date, trainer, shift?
← Select a scenario and run the engine
Core Algorithm — Gate-Based Failure Classification
systemic_detector.py — classification logic
def classify_qa_failure(criterion_id, failing_agents, eval_window, ctx):
    # GATE 1: System health
    events = ctx.get_system_events(eval_window)
    has_outage = any(e.type in ["OUTAGE", "LATENCY_SPIKE"] for e in events)
    n = len(failing_agents)
    rate = n / ctx.agents_evaluated_on(criterion_id, eval_window)
    if n >= 3 and has_outage:
        return Classification(type="SYSTEMIC", coaching_blocked=True)
    # GATE 2: Pattern width
    if n >= 3 and rate > 0.25:
        return Classification(type="SYSTEMIC", coaching_blocked=True)
    # GATE 5: Cohort pattern
    cohort = detect_cohort(failing_agents, ctx.roster)
    if cohort.significant:
        return Classification(type="COHORT")
    # All gates passed → individual
    return Classification(type="INDIVIDUAL", coaching_blocked=False)
Systemic gate must clear before individual scoring activates
⊘ BLOCKED — Systemic check required before individual analysis
Full Data (6 dimensions)
Partial Data (5 dimensions)
Adaptive Weights
When dimensions are missing, the engine redistributes weights and normalizes to 1.0. Scores shift. Rankings may change.
Systemic-First Gate
Engine refuses to produce scores until environmental factors are cleared. No scoring during outages or volume spikes.
Intervention Tiers
Monitor (≥70) → Coaching (50–69) → PIP (35–49) → Escalation (<35). Validated thresholds.
Architecture Decisions — Why This Design
Why Gates, Not ML
ML classifiers need labeled training data. Nobody labels QA failures as systemic vs individual. Gates produce those labels as a byproduct.
Why coaching_blocked
Without a hard gate, supervisors default to coaching. coaching_blocked = True removes that option when the problem isn't the person.
Deming's Insight
85% of quality problems originate in the system, not the worker. Contact centers have never systematically applied this. This engine does.
5
Classification Gates
p<.05
Cohort Detection
6+1
Failure Scenarios
6
Scoring Dimensions
4
Intervention Tiers
Demand Forecasting
Accuracy Workflow
Select a month and adjust parameters — watch the forecast decompose through seasonal + staffing layers
12-Month Seasonal Curve (Monthly Volume Index)
Forecast MonthJuly
Growth Rate+2%
Service Level Target80/120
Shrinkage28%
AHT Override (min)10.7
Current Headcount120
Weekly Volume
14,835
MVI 133.3 × 11,072 baseline
Peak Day (Mon)
2,621
DOW multiplier: 1.237
Required Agents (Peak Hr)
148
Erlang C @ 80/120 SL target
Required (Avg Hr)
112
Includes shrinkage adjustment
Staffing Gap
-28 agents
Peak hour deficit against current headcount
Multiplicative Decomposition
Baseline × Monthly Index × DOW Multiplier × Hourly Distribution. Three layers of seasonal adjustment before Erlang C runs.
Validated Model
Trained on 2024 data, predicted 2025: Pearson r=0.886, MAPE 6.2%. Peak and trough identified correctly.
Erlang C Staffing
Queueing theory calculates minimum agents for target SL at given arrival rate, handle time, and wait threshold.
20-Week Forecast vs. Actual — Toggle Drift to See Self-Correcting Detection
Forecast Actual
6.2%
Overall MAPE
5.8%
Weighted MAPE
+1.3%
Bias
0.8
Tracking Signal
⚠ DRIFT DETECTED: 4+ consecutive weeks above 15% MAPE. Model baseline requires recalibration.
Self-Correcting
Weekly workflow compares forecast to actual. When MAPE exceeds 15% for 4+ weeks, recalibration alert fires. The model watches itself.
Tracking Signal
Cumulative bias ÷ MAD. When it exceeds ±4.0, the model has systematic error — not random variance.
Production Safety Net
Most AI portfolios show a model built once. This shows it monitored continuously with automated drift detection.
104
Weeks Validated
1.34M
Calls Analyzed
0.886
Pearson r
6.2%
Monthly MAPE
12mo
ML Bootstrap