Abhiram Elangovan

Building Analytics Infrastructure | 5+ Years Experience | Gen AI Ready

Overview

This portfolio demonstrates my automation-first approach to building BI systems across three distinct contexts:

1. Scaling Enterprise Infrastructure - P&G Supply Chain Analytics

1. Building from Zero for Startups - Via.work BI Platform

1. Leveraging AI to Eliminate Manual Processes - GenAI-Powered Classification
   

Each project follows a consistent framework: Problem → Architecture → (Automation as needed) → Outcome

Project 1: Global Supply Chain Analytics Infrastructure

Client: Procter & Gamble (Fortune 30) | 2019-2021 (around 1,5 years project duration)

Core Challenge: Modernise a fragmented, 20+ day manual reporting process into an automated, scalable analytics platform serving 24 APAC markets and data across 28,000 stores

📊 The Problem

Business Context
P&G's legacy distribution tracking system was unable to support proactive decision-making:

• Data scattered across 6+ disparate systems (Azure, SQL, SAP, manual exports)

• Dashboard refresh took 20+ days, making insights obsolete before delivery

• No unified view across 24 countries, 20,000+ stores, 60M+ rows of store-SKU data

• 100+ Country and Category managers lacked self-service capabilities
  

Technical Debt

• Siloed data schemas with no harmonization

• Manual data pulls and transformations

• No security model for multi-market access

• Poor performance at scale
  

Stakeholder Impact
Business and IT Country Heads couldn't make timely supply chain optimization decisions, resulting in missed opportunities for distribution improvements and inventory management.

🏗️ The Architecture

Phase 1: MVP & Stakeholder Buy-in (Solo, 6 months)

┌─────────────────┐
│ Single Country  │ → Manual Data Pulls →  PowerBI Desktop              → Category Manager
│ (Thailand)      │                       (Data Model + Local Testing)
└─────────────────┘

Key Activities:

• Worked directly with Singapore Category Manager to understand pain points

• Built prototype dashboard with core Distribution KPIs

• Demonstrated value: Reduced their reporting time from 5 days to 2 hours

• Won stakeholder trust through iterative feedback and quick wins

Technologies Used:

• Power Query (data cleaning & transformation)

• DAX (custom measures and KPIs)

• Power BI Desktop (visualization)

Outcome:
Secured buy-in from Country Head to scale solution across entire APAC region

Phase 2: Scale & Automation (Led team of 3, 12 months)

┌─────────────┐     ┌──────────────┐     ┌─────────────┐     ┌──────────────┐
│ Data Sources│ →  │  ETL Layer   │ →  │   Semantic │ →   │  Power BI   │
│ (6 schemas) │    │              │    │    Layer   │     │   Premium   │
│             │    │              │    │  (Azure AAS)│    │   (P2 Tier)│
└─────────────┘     └──────────────┘     └─────────────┘     └──────────────┘
     │                    │                    │                    │
     ├─ Azure DB          ├─ Harmonization     ├─ DAX Measures      ├─ RLS (Country)
     ├─ SQL Server        ├─ Data Quality      ├─ Business Logic    ├─ Self-Service
     ├─ SAP               ├─ Transformation    ├─ Semantic Model    ├─ Incremental
     └─ Manual Exports                         └─ Single Source     └─   Refresh

Team Structure:

• Me (Lead Consultant): Solution architecture, stakeholder management, semantic layer design, Power BI development + 2 BI Developers

• 2 External Data Engineers: ETL pipeline development, data quality frameworks, SQL ETL experts

• Total Team: 5 people

Architectural Decisions:

1. Data Engineering Layer (Azure)

   • Automated data ingestion from 6 disparate sources

   • Built data quality validation framework

   • Implemented data harmonization logic

   • Created staging → bronze → silver → gold data pipeline

1. Semantic Layer (Azure Analysis Services)

   • Moved complex calculations from Power BI to backend

   • Designed star schema with fact and dimension tables

   • Centralized business logic for consistency

   • Optimized for query performance

1. Consumption Layer (Power BI Premium P2)

   • Deployed to enterprise-grade infrastructure

   • Implemented Row-Level Security (RLS) at country level

   • Used bookmarks for user-centric page navigation

   • Optimized visuals for 100+ concurrent users

Refresh Schedule:

• Incremental Refresh (Daily): New/updated data only, refreshes overnight (~4 hours)

• Full Refresh (Weekly): Entire dataset, runs over weekend (~18 hours)

🤖 The Automation Factor

What We Automated:

1. Data Ingestion

   • Eliminated manual exports from SAP, SQL, Azure

   • Built scheduled Azure jobs for automatic data pulls

   • Implemented retry logic and error handling

1. Data Quality Checks

   • Automated validation rules for data completeness

   • Anomaly detection for unusual data patterns

   • Alerting system for data quality issues

1. Transformation & Harmonization

   • Standardized 6 different schemas into unified model

   • Automated currency conversions and unit standardizations

   • Built reusable transformation templates

1. Dashboard Refresh

   • Moved from manual → scheduled incremental refresh

   • Implemented intelligent caching strategies

   • Optimized refresh for performance (incremental vs. full)

1. Distribution & Access

   • Self-service model: Users access dashboards directly

   • Automated RLS provisioning based on user roles

   • Eliminated manual report distribution

Why Automation Was Critical:

• Scale: 24 markets × multiple categories = impossible to maintain manually

• Timeliness: Business decisions needed current data, not 20-day-old data

• Consistency: Automated pipelines ensure standardized calculations across markets

• Reliability: Reduced human error and dependency on individual analysts

📈 The Outcome

Client Feedback:

Business Impact:

Specific Use Cases Enabled:

1. Out-of-Stock Prevention: Identified distribution gaps before they became stockouts

1. Retail Execution: Tracked in-store visibility and shelf placement effectiveness

1. Promo Performance: Measured real-time impact of marketing initiatives

1. Market Comparison: Benchmarked performance across countries

Financial Impact
Project success led to $50,000 additional revenue for my consulting firm through contract expansion

🎯 Key Learnings

1. Start Small, Scale Fast: Phase 1 MVP earned stakeholder trust, making Phase 2 scaling possible

1. Move Compute to Backend: Shifting calculations from Power BI to semantic layer dramatically improved performance

1. Design for Self-Service: Empowering users to explore data reduces ad-hoc requests and scales better

1. Automation Requires Investment: Building robust pipelines takes time upfront but pays massive dividends at scale

1. Data Quality is Non-Negotiable: Automated quality checks maintain stakeholder trust in self-service environment

Project 2: Startup BI Infrastructure from Ground Up

Client: Via.work (Series B Startup, now part of Justworks) | July 2023 - Jan 2024

Core Challenge: Build automated financial reporting and cross-functional analytics infrastructure during critical acquisition phase

📊 The Problem

Business Context:
Via.work, a Series B HR/payroll startup, was undergoing acquisition by Justworks:

• CEOs needed automated monthly financial updates for investors and board

• No unified view across Sales, Operations, and Finance verticals

• Manual consolidation of CRM data, revenue metrics, and operational KPIs

• Teams working in silos with no cross-functional visibility

Technical Gap

• No existing BI infrastructure or analytics team

• Data scattered across Hubspot CRM, Stripe (payments), spreadsheets

• Manual monthly report generation taking 3-4 days

• High risk of errors in critical acquisition-period reporting

Stakeholder Impact
CEOs spending valuable time consolidating data manually instead of focusing on acquisition strategy; teams unable to understand how their actions impacted other verticals or company bottom line.

🏗️ The Architecture

Approach: Build minimum viable analytics infrastructure that grows with the company

┌──────────────────┐     ┌─────────────────┐     ┌──────────────────┐
│  Data Sources   │ →   │  Transformation │ →  │  Consumption   │
│                 │     │  & Semantic     │    │  Layer         │
│                 │     │  Layer          │    │                │
└──────────────────┘     └─────────────────┘     └──────────────────┘
     │                          │                        │
     ├─ Hubspot CRM            ├─ Power Query           ├─ Executive
     │  (Sales data)           │  (ETL logic)           │  Dashboard
     │                         │                        │  (Investor
     ├─ Stripe + Internal SQL  ├─ DAX Semantic          │   Updates)
     │  (Operational)          │  Layer                 │
     │                         │  (KPIs & Metrics)      ├─ Sales
     ├─ Google Sheets          │                        │  Dashboard
      (Revenue Data)           ├─ Data Validation       │  (Pipeline)
                               │  Rules                 │
                               │                        └─ Operations
                               └─ Automated Refresh         Dashboard
                                  Schedule                (Efficiency)

Built from Scratch: Since no BI infrastructure existed, I architected the complete solution:

1. Data Integration Layer

   • Connected to Hubspot CRM via Python Pulls

   • Integrated SQL operations data

   • Consolidated operational data from Google Sheets and internal systems

1. Semantic Layer (Power BI)

   • Designed data model connecting all sources

   • Created business logic for key metrics:

     • Financial: MRR (Monthly Recurring Revenue), ARR, Churn Rate, ARPU (Average revenue per Unit)

     • Sales: Pipeline velocity, Lead-to-Customer conversion, CAC (Customer Acquisition Cost)

     • Operations: Customer onboarding time, Support ticket resolution

   • Built calculated measures translating technical data into business KPIs

1. Dashboard Layer

   • Executive Dashboard: Monthly board updates with financial health metrics

   • Sales Dashboard: Pipeline tracking, conversion funnels, win rates

   • Operations Dashboard: Operational efficiency, customer success metrics

   • Cross-Functional View: How each vertical impacts company performance

🤖 The Automation Factor

What We Automated:

1. Financial Reporting (Built on Google Sheets

1. Data Validation

1. Cross-Functional Insights

1. Report Distribution
   

Why Automation Was Critical:

• Acquisition Timing: Every day counted; CEOs needed to focus on deal, not data

• Accuracy: Manual consolidation risked errors in critical investor communications

• Agility: Business moving fast; needed real-time insights to adapt strategy

• Scalability: Built foundation for post-acquisition analytics needs

📈 The Outcome:

Prototype Reports:

Decision-Making Transformation

Before:

• Teams operated in silos without visibility

• Financial reporting was reactive and manual

• No understanding of cross-functional impact

After:

• Teams understood how their work impacted company goals

• Proactive identification of trends and issues

• Data-driven conversations across all verticals

• Reliable, automated financial reporting built investor confidence

• Clean data room accelerated due diligence process

Specific Use Cases Enabled:

1. Investor Communication: Real-time financial health metrics for board meetings

1. Sales Strategy: Pipeline visibility enabled better forecasting and Operations Team resource allocation

1. Strategic Planning: Cross-functional data informed product and go-to-market decisions

🛠️ Technical Stack

Data Sources:

• Hubspot CRM (API integration)

• SQL (client data)

• Google Sheets (operational + Revenue data)

• Internal databases

ETL & Transformation:

• Power Query (data transformation)

• Custom API connectors + Python

Semantic Layer:

• Power BI data modeling

• DAX (calculated measures and KPIs)

• Relationship management

Visualization:

• Power BI Desktop (development)

• Power BI Service (deployment)

• Shared workspaces for collaboration

Key Features:

• Self-service dashboards

• Cross-source data validation

• Mobile-responsive design

🎯 Key Learnings

1. Build for Self-Service Early: Empowering teams to access their own data reduced my bottleneck as solo BI person

1. Start with CEO Pain Points: Solving executive reporting needs first secured buy-in for broader analytics investment

1. Semantic Layer is Critical: Centralizing business logic in data model ensured consistency across all dashboards

1. Automation Builds Trust: Reliable, automated reports during acquisition built confidence with investors

1. Cross-Functional Visibility Drives Collaboration: When teams see how they impact each other, they work together better

Project 3: AI-Powered Multilingual Classification System (AI Project)

Client: Luxembourg Government | Master's Thesis, 2024-2025

Core Challenge: Automate manual 3-language job classification while ensuring GDPR compliance and mitigating algorithmic bias

📊 The Problem

Business Context
Luxembourg's labor market operates in three official languages (French, German, English):

• Government needed to classify job advertisements into standardized skill categories

• Manual process required analysts to read and categorize each job ad

• Language barriers made consistent classification difficult and biased

• Scale problem: Thousands of job ads monthly, growing backlog

Technical Challenges

• Multilingual NLP: Models trained on one language perform poorly on others

• GDPR Compliance: Cannot send sensitive job data to external LLM APIs

• Algorithmic Bias: Risk of model favoring certain languages or demographic groups

• Data Scarcity: Limited labeled training data across all three languages

Stakeholder Impact
Labor market analysts spending weeks manually classifying jobs instead of analyzing trends; inconsistent categorization across languages undermining data quality for policy decisions.

🏗️ The Architecture

Approach: Build GDPR-compliant, bias-aware AI pipeline for automated classification

┌─────────────────┐      ┌──────────────────┐       ┌─────────────────┐      ┌──────────────┐
│  Job Ads        │ →   │  Translation    │ →    │  Classification │ →   │  Validation │
│  (FR/DE/EN)     │     │  Layer          │      │  Layer          │     │  & Output   │
│                 │     │  (Self-hosted   │      │  (mBERT)        │     │             │
└─────────────────┘      │   LLM)          │      │                 │      └──────────────┘
     │                  └──────────────────┘       └─────────────────┘           │
     │                          │                        │                    │
     ├─ French ads              ├─ Offline hosting      ├─ Multilingual       ├─ Bias
     ├─ German ads              │  (GDPR compliant)     │  BERT model         │  metrics
     └─ English ads             │                       │                     │
                                ├─ GenAI-powered        ├─ Fine-tuned on      └─ Skill
                                │  translation          │  job data            categories
                                │                       │                      
                                └─ Data stays on-prem   └─ Synthetic data
                                                           for bias reduction

System Components:

1. Translation Layer (Self-Hosted LLM)

   • Deployed open-source LLM locally for GDPR compliance

   • Used for translating job ads to common language

   • Offline workflow: No data leaves government infrastructure

   • GenAI-powered semantic understanding beyond word-for-word translation

1. Classification Layer (mBERT)

   • Multilingual BERT model (mBERT) as core classifier

   • Fine-tuned on labeled job advertisement data

   • Trained to classify into standardized skill taxonomy

1. Bias Mitigation Framework

   • Generated synthetic training data using LLMs

   • Balanced representation across languages

   • Statistical testing for fairness across demographic groups

   • Monitoring dashboard for ongoing bias detection

1. Validation & Quality Assurance

   • Human-in-the-loop validation for edge cases

   • Confidence scoring for each classification

   • Feedback mechanism for continuous improvement

Why Automation + AI Was Critical:

• Scale: Manual process couldn't keep pace with job ad volume

• Consistency: Eliminated subjective interpretation differences between analysts

• Language Equality: Ensured no language disadvantaged in classification

• Efficiency: Freed analysts to focus on policy analysis vs. manual categorization

• Privacy: Self-hosted solution maintained GDPR compliance

AI/GenAI Usage:

📈 The Outcome

Business Impact

Research Contributions

1. GDPR-Compliant AI Architecture

   • Demonstrated how to deploy GenAI/LLMs offline for sensitive government data

   • Provided framework for EU organizations to leverage AI while maintaining compliance

   • Advisory report submitted to Luxembourg and Dutch governments

1. Multilingual Bias Mitigation

   • Analyzed algorithmic bias across three languages

   • Developed methodology for ensuring language equality in NLP systems

   • Published techniques for synthetic data generation to reduce bias

1. Production-Ready POC

   • Delivered working proof-of-concept system

   • Documented deployment architecture for government IT teams

   • Created monitoring dashboard for ongoing quality assurance

Specific Use Cases Enabled:

1. Labor Market Analysis: Faster, more consistent skill demand tracking

1. Policy Making: Reliable data on skill gaps for education policy

1. Job Seeker Support: Better matching between job ads and training programs

🛠️ Technical Stack

Data & Infrastructure:

• Python (data processing & model training)

• Self-hosted LLM (translation)

• On-premise compute (GDPR compliance)

Machine Learning:

• mBERT (Multilingual BERT) - core classifier

• Hugging Face Transformers library

• PyTorch (model training framework)

• Scikit-learn (evaluation metrics)

GenAI/LLM Components:

• Open-source LLM (offline translation)

• Fine-tuned LLM for synthetic data generation

• Prompt engineering for data augmentation

Bias & Quality:

• Statistical significance testing

• Fairness metrics (demographic parity, equal opportunity)

• Custom monitoring dashboard

• Human-in-the-loop validation interface

Key Innovations:

• GDPR-compliant LLM hosting architecture

• Synthetic data generation for bias reduction

• Multilingual model evaluation framework

• Production deployment guidelines for government use

🎯 Key Learnings

1. GDPR and AI Can Coexist: Self-hosted LLMs enable GenAI benefits while maintaining compliance

1. Bias is a Feature, Not a Bug: Proactive bias detection must be built into AI systems from day one

1. Synthetic Data is Powerful: LLM-generated training data can reduce real-world bias when done carefully

1. Multilingual NLP is Hard: Language equality requires intentional design and continuous monitoring

1. Government AI Needs Different Approach: Privacy, explainability, and fairness take precedence over raw performance

Conclusion

These three projects demonstrate my core philosophy: Build analytics systems that eliminate manual work through intelligent automation.

Common Themes Across Projects:

1. Automation-First Mindset

   • P&G: 20+ days → 24 hours through automated pipelines

   • Via.work: 3-4 days → 30 minutes for financial reporting

   • GenAI Prototype: 10-20 ads/day → 1000+/hour classification

1. Semantic Layer Design

   • P&G: Azure AAS semantic layer for 60M+ rows

   • Via.work: Power BI data model connecting CRM + finance + ops

1. Self Serve Analytics as a Product

   • P&G: 100+ users with RLS access

   • Via.work: 10+ managers exploring data vs. CEO-only reports

   • GenAI: Analysts focus on insights vs. manual categorization

1. Scale Through Architecture

   • P&G: 1 country → 24 countries, 1 user → 100+ users

   • Via.work: Built for growth from acquisition through scale

   • GenAI: 3 languages, production-ready for thousands of job ads

1. Data Quality as Foundation

   • P&G: Automated validation rules and anomaly detection

   • Via.work: Cross-source reconciliation and validation

   • GenAI: Bias monitoring and confidence scoring

Thank you!

Abhiram Elangovan
📧 mail@abhiram.me
📱 +31-657881547
🔗 linkedin.com/in/elanabhi

All sensitive data has been redacted or anonymized. Happy to discuss any project in detail during the interview process.