Alexander AI HR Assistant

Project Documentation

Problem

New employees face fragmented onboarding with information scattered across systems like SharePoint and Confluence. HR teams spend excessive time answering repetitive questions instead of strategic work. 86% of HR professionals expect onboarding to help new hires feel at ease, but 49% report inadequate monitoring and 48% cite inconsistent application across organizations.

Solution

Alexander: AI-powered onboarding assistant providing instant answers to policy questions, guiding new employees through company documentation, and reducing HR workload through automated knowledge retrieval.

My Role

Product Designer & Technical Architecture Lead

Tech Stack

• Python
• FastAPI
• OpenAI
• Vector DB
• Neo4j
• PostgreSQL

Overview

Alexander is an AI-powered HR onboarding assistant designed for LSE's IT Service Innovation course. The project addresses how new employees navigate complex organizational knowledge during their critical first weeks, when information overload and hesitation to ask questions creates productivity gaps.

Developed as an end-to-end product design covering technical architecture, user experience, system design, and implementation strategy for enterprise deployment.

Challenge Context

• Course: IT Service Innovation (LSE)
• Brief: Design AI solution addressing organizational challenge with practical implementation plan
• Focus: Multi-model architecture, data governance, and enterprise integration
• Team: 4 members

Technical Architecture Design

Three-Model AI System

Model 1: NLP Engine (BERT)

• Intent recognition and entity extraction
• Processes user queries to understand context
• Fine-tuned on HR-specific language patterns
• Handles ambiguous questions through contextual understanding

Model 2: Machine Learning Recommendations (XGBoost)

• Collaborative filtering based on user behavior
• Content-based filtering matching user profiles to relevant documents
• Learns from feedback to improve suggestion accuracy
• Predicts next-likely questions based on role and department

Model 3: Knowledge Graph (Neo4j)

• Organizes company documents, policies, and relationships
• Maps connections between entities (employees, departments, policies)
• Enables semantic search beyond keyword matching
• Provides context-aware answers based on user's role and access level

System Design Principles

User-Centric Design

• Clean chat interface with document suggestions
• Home screen showing frequently asked questions and relevant materials
• Role-based access control ensuring appropriate information visibility
• Mobile-first responsive design for accessibility

HR & IT Oversight

• Admin dashboard for content management
• Usage analytics showing common questions and knowledge gaps
• Feedback loop for continuous model improvement
• Four-tier confidentiality system for document access control

Scalability & Adaptability

• Cloud-based architecture (AWS S3, Lambda, EC2)
• API-first design for integration with existing systems
• Serverless functions for dynamic scaling
• Modular components allowing independent updates

Security

• End-to-end encryption for data transmission
• GDPR-compliant data handling
• Role-based authentication
• Audit trails for compliance monitoring

Data Pipeline Architecture

Stage 1: Data Collection

• API integration with knowledge management systems
• Manual upload with OCR validation for documents
• Central PostgreSQL database with real-time validation

Stage 2: Preprocessing

• Automated deduplication and quality checks
• Feature engineering (shelf life, categorization, geospatial)
• Natural language processing for text extraction

Stage 3: Knowledge Graph Construction

• Entity relationship mapping
• Graph embeddings for semantic search
• Vector representations using ElasticSearch with k-NN

Stage 4: Response Generation

• Multi-model orchestration selecting optimal answer source
• BERT generates natural language responses
• Citations to source documents with access verification

Stage 5: Continuous Learning

• User feedback integration
• Performance monitoring and model retraining
• A/B testing for algorithm improvements

My Contributions

• Designed complete system architecture including three-model pipeline
• Created user journeys for employees, HR managers, and administrators
• Developed technical specifications for each component interaction
• Designed role-based access control framework
• Built evaluation framework for measuring system effectiveness
• Coordinated 4-person team across design and research phases

Innovation Highlights

• Multi-model architecture providing comprehensive knowledge retrieval
• Context-aware responses adapting to user's role and department
• Proactive suggestion system predicting information needs
• Privacy-first design with granular access controls

Implementation Strategy

• SaaS deployment model for enterprise clients
• Initial training on company-specific documentation
• Phased rollout starting with pilot departments
• Integration with existing tools (Slack, Microsoft 365, Jira)

Lessons Learned

System Design: Multi-model approaches provide better coverage than single solutions but require careful orchestration to manage latency

User Experience: Change management crucial for adoption—employees need confidence that AI answers are trustworthy

Data Governance: Role-based access and audit trails must be designed from the start, not added later

Scalability: Early architecture decisions around cloud infrastructure and API design determine future flexibility

Additional Documentation