🎙️ Voice AI Analytics

Role: Analytics Engineer Stack: Docker · PostgreSQL · dbt · Python · Metabase · VS Code

---

📌 Project Overview

This project analyzes Voice AI interactions to evaluate accessibility, efficiency, adoption, friction, and error reduction using real conversational data.

The solution is implemented using a modern analytics engineering stack, emphasizing:

• Reproducibility
• Data quality & governance
• Clear analytical reasoning
• Stakeholder-ready insights

The final output supports:

• KPI tracking
• Root-cause analysis of friction and abandonment
• Error-reduction measurement
• Trustworthy dashboards via Metabase

---

🧱 Full Tech Stack

Layer	Technology	Purpose
Orchestration	Docker / Docker Compose	Reproducible local environment
Database	PostgreSQL	Raw + analytics warehouse
Transformation	dbt	Staging, marts, testing, governance
Analysis	Python (Pandas, SQLAlchemy, sklearn)	Statistical analysis & modeling
BI / Visualization	Metabase	Stakeholder dashboards
Dev Environment	VS Code (Dev Containers)	Seamless containerized development

---

🏗️ Architecture Overview

CSV Data
  ↓
PostgreSQL (raw tables)
  ↓
dbt Staging Models (stg_*)
  ↓
dbt Fact Models (fact_voice_ai_sessions)
  ↓
Metabase Dashboards
  ↓
Stakeholder Insights

Design principles:

• One fact table = one grain (session-level)
• No business logic in dashboards
• All metrics validated before exposure
• Python used only where SQL is insufficient

---

🚀 Environment Setup (Step-by-Step)

1️⃣ Prerequisites

• Docker & Docker Compose
• VS Code + Dev Containers extension

---

2️⃣ Clone Repository

git clone <repo-url>
cd <repo-name>

---

3️⃣ Build & Start Containers

docker compose up --build

This starts:

• postgres (persistent volume)
• analytics (Python + dbt)
• metabase

---

4️⃣ Attach VS Code to Container

In VS Code:

Command Palette → Dev Containers: Attach to Running Container
→ analytics

Your local workspace is now mounted into the container.

---

5️⃣ Load Data into PostgreSQL

Data is loaded once and persists via Docker volumes.

docker exec -it postgres psql -U analyst -d irembo

---

6️⃣ dbt Setup

cd /workspace/dbt
dbt deps
dbt debug

---

7️⃣ Run Transformations & Tests

dbt build

This will:

• Build staging models
• Build fact models
• Run data quality tests

---

8️⃣ Access Metabase

http://localhost:3000

Connect Metabase to:

• Host: postgres
• Database: irembo
• Schema: analytics

---

📂 Repository Structure

.
├── docker-compose.yml
├── Dockerfile
├── requirements.txt
├── data/
|     └── raw/
|          └── all.csv
├── scripts/
|       └── load_csvs.py
├── dbt/
│   ├── dbt_project.yml
│   ├── packages.yml
│   ├── models/
│   │   ├── staging/
│   │   └── marts/ 
│   └── tests/
|        └── tests.sql
├── notebooks/
│   └── 01_friction_analysis.ipynb
└── README.md

---

🧠 Walkthrough & Approach

---

Part 1: Analytics Design & Monitoring (KPIs)

🎯 Goals & KPIs

Accessibility / Inclusivity

• Completion rate for first-time users
• Rural vs urban completion gap
• Average ASR confidence by user type

Efficiency

• Average turns per session
• Average session duration
• Error rate per session

Adoption

• Share of sessions using Voice AI
• Repeat usage rate
• Voice vs non-voice completion delta

Approach: KPIs are session-level metrics aggregated from turn-level data and materialized in fact_voice_ai_sessions.

---

Part 2: Data Modeling

Core Fact Table

fact_voice_ai_sessions

• Grain: 1 row per session

• Source: voice_turns only (applications excluded after validation)

• Metrics:

  • Error proportions
  • ASR & intent confidence
  • Turn counts
  • Final outcome

Why: Sessions are the natural decision unit for Voice AI effectiveness.

---

Part 3: Insight Generation

Key Findings (Stakeholder Summary)

1. Primary friction driver is ASR performance, not conversation length.
2. Users abandon early, indicating initial recognition issues.
3. Voice AI performs better for first-time digital users than non-voice channels.
4. Rural users benefit disproportionately from Voice AI accessibility.

Tools Used:

• SQL → aggregations & KPIs
• Python → Logistic Regression, segmentation, statistical validation

---

Part 4: Impact & Error Reduction

Defining an Error

An error is any turn marked as:

• misunderstanding
• silence
• or repeated intent failures

Baseline Error Rate

error_turns / total_turns

Measuring Improvement

• Compare pre/post model deployments
• Control for user mix & session length
• Track statistically significant deltas

Avoiding Misleading Conclusions

• Segment by channel & user type
• Avoid raw averages
• Require confidence intervals

Twist: Logistic Regression Insight

Modeling non-completion revealed:

• ASR confidence as the strongest abandonment predictor
• Errors occur early, not due to long sessions

➡️ Strategy shift: Fix ASR first, not dialogue flow.

---

Part 5: Data Quality & Governance

dbt Tests Implemented

• Not null & uniqueness

• Relationship tests

• Accepted ranges (0–1)

• Business logic tests:

  • Error proportions ≤ 100%
  • Sessions must have turns

PII Protection

• No raw audio
• User IDs anonymized
• Aggregated reporting only

Metabase only queries certified dbt models.

---

Part 6: Production Readiness

What Makes This Production-Grade

• Deterministic builds
• Persistent volumes
• Versioned transformations
• Automated data quality enforcement
• BI governance boundary

---

🏆 Final Notes

This project was designed as:

• A real analytics platform, not a one-off analysis
• A showcase of 6+ years Analytics Engineering maturity
• A system that scales beyond the assignment

Key philosophy: If Metabase can see it, it has already been tested.

---

✅ Next Steps (Optional Enhancements)

• CI pipeline for dbt tests
• Snapshotting for longitudinal analysis
• Feature store for ML reuse

---