Skip to content

wiwidcok/nuxhub-simulate

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

NuxHub Simulate — AI Sales Pipeline Simulation Engine

Generate random lead profiles, run full conversation simulations through your NuxHub AI pipeline, QA every artifact, and measure how effectively your AI agents move leads toward conversion.

License: MIT Hermes Agent NuxHub

What is NuxHub Simulate?

NuxHub Simulate is a skill pack for Hermes Agent that turns your NuxHub AI sales pipeline into a testable, measurable system. Instead of waiting for real leads to test your AI agents, NuxHub Simulate generates realistic random lead profiles and runs them through your entire pipeline — from first message to conversion — then analyzes how well your AI responds at every step.

Think of it as a wind tunnel for your AI sales agents.

The Problem It Solves

You built an AI sales pipeline with multiple agents (router, discovery, qualification, objection, closing). But how do you know:

  • Is the AI responding with the right tone for each personality type?
  • Is discovery collecting enough information before pushing offers?
  • Are objections being handled effectively or defensively?
  • Is the lead actually progressing toward conversion or getting stuck?
  • Are edge cases (scammers, angry users, elderly users) handled correctly?

NuxHub Simulate answers all of these by running synthetic conversations through your real pipeline and analyzing every turn.


Key Features

1. Random Profile Generation

Generate 10+ realistic lead profiles with diverse characteristics:

  • 4 personality types: Analytical, Driver, Amiable, Expressive
  • 8+ business types: Beauty clinics, auto repair shops, restaurants, online stores, pharmacies, resellers, wellness agencies, personal consumers
  • Variable BANT scores: Hot (80+), Warm (50-80), Cold (<50)
  • 8 intent levels: From "unknown" to "transaction_ready"
  • Edge cases (~20% of profiles): Scammers, toxic users, tech-illiterate users, enterprise leads, ghost leads, off-topic

2. Full Pipeline Simulation

Each profile runs through your complete AI pipeline as a multi-turn conversation:

  • Profile acts as the "user" sending messages
  • Your NuxHub orchestrator processes each message (spam gate → escalation → router → specialist → judge → response)
  • AI response is captured and analyzed
  • Profile generates a contextual reply based on personality type
  • Loops until terminal state: Conversion, Follow-up, Blacklist, Escalation, or Timeout

3. QA + Fixer Loop

Every artifact produced by your specialist agents is quality-checked:

  • Judge verifies schema compliance, data completeness, confidence accuracy, objective alignment
  • If an artifact fails QA, the fixer loop retries the specialist (max 3 attempts) with the error report
  • Fixer logs are stored for debugging

4. AI Response Quality Analysis

The core analysis focuses on how the AI responds to users:

  • Tone Match Rate: Does the AI adapt its tone to the user's personality type?
  • Question Relevance: Are the AI's questions relevant and progressive?
  • Empathy Score: Does the AI show understanding or respond robotically?
  • Pushiness Score: Is the AI pushing offers too early? (50 = ideal)
  • Objection Handling: Does the AI address root causes or get defensive?
  • Progress Rate: Is the conversation moving forward each turn?

5. Sales Progression Tracking

Track how leads move through the sales funnel turn by turn:

  • Intent level progression (unknown → curious → ... → transaction_ready)
  • BANT score progression per turn
  • Conversion readiness score (0-100) per turn
  • Stage reached: discovery → qualification → offering → closing → conversion
  • Next-level blocker identification

6. Obstacle & Bottleneck Detection

Every simulation logs obstacles with severity, root cause, and recommendations:

  • state_stuck — AI loops without progress
  • offer_premature — AI offers before discovery is complete
  • personality_misread — AI uses wrong tone for personality type
  • false_conversion — AI marks conversion when readiness < 90
  • spam_misclassification — Spam gate false positive/negative
  • escalation_miss — Escalation trigger not detected
  • objection_unresolved — Objection not resolved within max turns
  • And more...

7. Reports

Generate visual HTML + PDF reports with:

  • Executive summary with conversion rate and verdict
  • State reach waterfall (NEW_LEAD → CONVERSION)
  • AI response quality metrics dashboard
  • Sales progression analysis
  • Top obstacles and bottlenecks (sorted by frequency and impact)
  • Personality type breakdown
  • Edge case handling results
  • QA metrics (pass rate, fixer triggered, schema compliance)
  • Actionable recommendations for pipeline improvement
  • Per-profile detailed appendix

Outcomes for Users

When you install and run NuxHub Simulate, you get:

Outcome Description
Conversion Rate Measurement Know exactly what percentage of leads your AI pipeline successfully converts
AI Response Quality Scores Quantitative scores for tone, empathy, pushiness, question relevance — per turn and aggregated
Bottleneck Identification Know exactly which agent/state is slowing down or blocking lead progression
Personality Adaptation Audit Verify your AI adapts its tone correctly for analytical, driver, amiable, and expressive users
Edge Case Validation Confirm your spam gate catches scammers, your escalation gate catches enterprise/legal/complaint leads, and your AI handles tech-illiterate and elderly users with patience
Schema Compliance Report Verify every agent produces artifacts that comply with your schema definitions
Fixer Loop Effectiveness Track how often the fixer loop is triggered and how often it resolves issues
Sales Funnel Leakage Points Identify exactly where leads drop off in the journey from discovery to conversion
Actionable Recommendations Prioritized list of fixes to improve your pipeline's conversion rate
Reproducible Test Suite Profiles are stored as YAML files — re-run the same scenarios after fixes to verify improvement

Installation

Prerequisites

  • Hermes Agent installed and configured
  • A NuxHub AI project with skill packs (orchestrator, router, discovery, qualification, objection, closing, judge, follow_up)
  • Shared configs: personalities.yaml, products.yaml, goals.yaml, intent_ladder.yaml, spam_gate.yaml, state_machine.yaml, artifact_schemas.yaml

Install the Skill Pack

# Clone the repo
git clone https://github.com/wiwidcok/nuxhub-simulate.git

# Copy skill to Hermes skills directory
cp -r nuxhub-simulate/skill/nuxhub-simulate ~/.hermes/skills/nuxhub/

# Verify installation
# In Hermes, the skill should appear in skills list

Set Up a Simulation Project

# Create simulation folder (FHS standard: state data in /var/lib)
mkdir -p <simulation-data-path>/{your-project}-simulation/{profiles,artifacts,qa,fixer,analysis,reports,docs}

# Copy config template and edit to match your project
cp nuxhub-simulate/examples/config.yaml <simulation-data-path>/{your-project}-simulation/config.yaml

# Edit config.yaml to point to your project's shared configs
vim <simulation-data-path>/{your-project}-simulation/config.yaml

Usage

Basic Simulation (10 profiles)

simulate: {"project": "nuxhub-alpha-propolis", "count": 10}

Custom Count

simulate: {"project": "nuxhub-alpha-propolis", "count": 20}

Specific Phase Only

simulate: {"project": "nuxhub-alpha-propolis", "count": 10, "phase": "discovery"}

Parallel Simulations

simulate: {"project": "nuxhub-alpha-propolis", "count": 10, "parallel": 5}

Read Results

# Aggregate report
cat <simulation-data-path>/nuxhub-alpha-propolis-simulation/analysis/aggregate-report.yaml

# Per-profile analysis
cat <simulation-data-path>/nuxhub-alpha-propolis-simulation/analysis/sim-001-analysis.yaml

# HTML report
open <simulation-data-path>/nuxhub-alpha-propolis-simulation/reports/sim-latest.html

How It Works

┌──────────────────────────────────────────────────────────────┐
│                    NuxHub Simulate Engine                      │
├──────────────────────────────────────────────────────────────┤
│                                                              │
│  1. LOAD CONFIG                                              │
│     Read project's shared configs (personalities, products,   │
│     goals, state_machine, schemas)                            │
│                                                              │
│  2. GENERATE PROFILES (min 10)                                │
│     ┌────────────────────────────────────────┐               │
│     │ 8 normal profiles (2 per personality)  │               │
│     │ 2 edge cases (scam, tech-illiterate, etc.)│               │
│     └────────────────────────────────────────┘               │
│                                                              │
│  3. RUN SIMULATIONS (parallel via delegate_task)             │
│     ┌─────────────┐  ┌─────────────┐  ┌─────────────┐        │
│     │  Profile 1  │  │  Profile 2  │  │  Profile 3  │  ...   │
│     │             │  │             │  │             │        │
│     │ Turn 1:     │  │ Turn 1:     │  │ Turn 1:     │        │
│     │ user msg    │  │ user msg    │  │ user msg    │        │
│     │     ↓       │  │     ↓       │  │     ↓       │        │
│     │ Pipeline    │  │ Pipeline    │  │ Pipeline    │        │
│     │ (spam→route │  │ (spam→route │  │ (spam→route │        │
│     │  →specialist│  │  →specialist│  │  →specialist│        │
│     │  →judge)    │  │  →judge)    │  │  →judge)    │        │
│     │     ↓       │  │     ↓       │  │     ↓       │        │
│     │ AI response │  │ AI response │  │ AI response │        │
│     │     ↓       │  │     ↓       │  │     ↓       │        │
│     │ QA + Fixer  │  │ QA + Fixer  │  │ QA + Fixer  │        │
│     │     ↓       │  │     ↓       │  │     ↓       │        │
│     │ Turn 2...   │  │ Turn 2...   │  │ Turn 2...   │        │
│     │     ↓       │  │     ↓       │  │     ↓       │        │
│     │ Terminal    │  │ Terminal    │  │ Terminal    │        │
│     └─────────────┘  └─────────────┘  └─────────────┘        │
│                                                              │
│  4. ANALYZE                                                   │
│     Per-profile: state progression, AI quality, obstacles    │
│     Aggregate: conversion rate, bottlenecks, recommendations │
│                                                              │
│  5. REPORT                                                   │
│     HTML + PDF with visualizations and recommendations        │
│                                                              │
└──────────────────────────────────────────────────────────────┘

Project Structure

nuxhub-simulate/
├── README.md                    # This file
├── LICENSE                      # MIT License
├── skill/
│   └── nuxhub-simulate/
│       └── SKILL.md             # Skill pack instructions for Hermes Agent
├── examples/
│   ├── config.yaml              # Example simulation config
│   ├── profile-001.yaml         # Example: analytical, hot lead
│   ├── profile-002.yaml         # Example: driver, hot lead
│   ├── profile-003.yaml         # Example: amiable, warm lead
│   ├── profile-004.yaml         # Example: expressive, warm lead
│   ├── profile-005.yaml         # Example: analytical, warm lead
│   ├── profile-006.yaml         # Example: driver, cold lead
│   ├── profile-007.yaml         # Example: amiable, edge case: tech-illiterate
│   ├── profile-008.yaml         # Example: edge case: scammer
│   ├── profile-009.yaml         # Example: edge case: enterprise escalation
│   └── profile-010.yaml         # Example: expressive, warm lead
└── docs/
    ├── SIMULATION-GUIDE.md      # How to run simulations
    ├── PROFILE-TYPES.md         # Profile type reference
    └── ANALYSIS-METRICS.md      # Metrics reference

Profile Type Distribution

Per 10 profiles generated:

Category Type Count Purpose
Normal Analytical 2 Test data-driven responses
Normal Driver 2 Test fast, to-the-point responses
Normal Amiable 2 Test empathetic, relationship-building responses
Normal Expressive 2 Test enthusiastic, vision-driven responses
Edge Case Scam 1 Test spam gate blacklist
Edge Case Enterprise/Toxic/Tech-illiterate/Ghost 1 Test escalation/gate handling

Analysis Metrics

AI Response Quality

Metric Range Ideal
Tone Match Rate 0-100% >85%
Question Relevance 0-100 >80
Empathy Score 0-100 >70
Pushiness Score 0-100 50 (balanced)
Objection Handling 0-100 >75

Sales Progression

Metric Description
Intent Progression Track intent level per turn (0-8 scale)
BANT Progression Track BANT score per turn (0-100)
Readiness Progression Track conversion readiness per turn (0-100)
Stage Reached Highest stage: discovery → qualification → offering → closing → conversion
Next Level Blocker What prevents the lead from advancing

Verdict System

Verdict Condition
Healthy Conversion rate >40%, no critical obstacles, schema compliance >95%
Needs Attention Conversion rate 20-40%, or critical obstacles present
Critical Conversion rate <20%, or multiple critical obstacles

Configuration

Edit config.yaml to customize your simulation:

project:
  name: "nuxhub-alpha-propolis"
  source_path: "<your-nuxhub-source-path>"
  shared_configs:
    personalities: "/path/to/personalities.yaml"
    products: "/path/to/products.yaml"
    goals: "/path/to/goals.yaml"
    # ... etc

simulation:
  default_count: 10           # minimum profiles per run
  max_count: 50               # hard cap
  parallel: 3                 # concurrent simulations
  max_turns_per_sim: 30       # hard cap turns per profile
  timeout_per_sim: 300        # seconds
  fixer_max_retry: 3          # per artifact

profile_distribution:
  normal:
    analytical: 2
    driver: 2
    amiable: 2
    expressive: 2
  edge_cases: 2

Use Cases

1. Pre-Launch Validation

Before deploying your AI sales pipeline, run 10-20 simulations to verify it handles different personality types, edge cases, and objection scenarios correctly.

2. Continuous Improvement

After fixing a bottleneck (e.g., "discovery agent asks too many questions"), re-run the same profiles to measure the improvement quantitatively.

3. A/B Testing Pipeline Changes

Change a shared config (e.g., update personalities.yaml tone rules), run simulations, compare conversion rates before and after.

4. Team Reporting

Generate HTML+PDF reports to share with stakeholders showing pipeline performance, conversion rates, and improvement recommendations.

5. Regression Testing

Keep a fixed set of profiles as your regression test suite. After any pipeline change, run them to verify no regressions.


Compatibility

  • Hermes Agent — This is a Hermes Agent skill pack. It requires Hermes Agent to run.
  • NuxHub AI — Designed for projects using the NuxHub AI architecture (state-based routing, multi-agent pipeline, artifact-based communication).
  • Any NuxHub Project — Generic by design. Create a new simulation folder + config for each NuxHub project you want to test.

License

MIT License — see LICENSE file for details.


Author

Yeremia Widya Parama (@wiwidcok)

Built with Hermes Agent by Nous Research.


Keywords

AI sales pipeline simulation, lead profile generation, conversation simulation, AI response quality analysis, sales progression tracking, multi-agent pipeline testing, BANT scoring simulation, personality-based AI testing, AI sales agent QA, conversion rate measurement, sales funnel bottleneck detection, NuxHub AI, Hermes Agent skill pack, AI lead nurturing validation, synthetic conversation testing, AI agent quality assurance.

About

AI Sales Pipeline Simulation Engine — Generate random lead profiles, run full conversation simulations through your NuxHub AI pipeline, QA every artifact with fixer loop, analyze AI response quality, and measure sales progression. Hermes Agent skill pack.

Topics

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors