Enterprise-Scale AI Survey Engine for HR SaaS

The client — a fast-scaling SaaS HR platform — had a clear objective: kill the static survey. Replace it with a multilingual AI engine that listens at scale, analyzes in real time, and delivers insights HR leaders can act on.

Their core platform already helped enterprises manage engagement, development, and retention through data. But static forms stalled momentum. They needed a system that could surface sentiment — and turn it into strategy.

We delivered an architecture built for scale and speed. AI handled sentiment analysis, flagged risk patterns, and generated reports the moment feedback landed. No delay. No guesswork. HR teams saw what mattered — and acted faster.

The new module positioned the client not as a company with surveys, but as a platform that decodes employee signals before they become attrition stats.

• Solution Architect (1)
• AI/ML Engineers (2)
• Backend Developers (3)
• Frontend Developers (2)
• Data Engineer (1)
• QA Engineers (2)
• DevOps Engineer (1)
• UI/UX Designer (1)
• Product Owner (Client-side)

• Outdated monolithic PHP survey engine, inflexible and regression-prone, unfit for real-time analytics.
• Low engagement and survey fatigue caused by static, non-contextual questionnaires.
• Slow feedback-to-action cycle, with up to a week for analysis, delaying response to emerging HR issues.
• Disconnected data silos where feedback, HRIS, and performance metrics are stored separately, blocking correlation.
• Integration requirements with Workday, BambooHR, SAP SuccessFactors, and SSO through Azure AD, Okta, and Google Workspace.
• High-load scalability requirements as campaigns generate tens of thousands of responses in hours, stressing compute and storage.
• Role-based access, encryption, anonymization, and audit logging are required for end-to-end security.

• Frontend: React 18, TypeScript, Material UI
• Backend: Node.js 20 (NestJS), GraphQL API
• Database: PostgreSQL 16, Redis for caching
• AI/ML: OpenAI GPT-4.5 API (custom fine-tuning), Hugging Face Transformers, spaCy for NLP preprocessing
• Data Processing & ETL: Apache Kafka, Python (Pandas, NumPy)
• Analytics & Visualization: Apache Superset, D3.js
• Cloud Infrastructure: AWS (ECS Fargate, S3, RDS, Lambda), Terraform for IaC
• Security: AWS KMS, JWT authentication, OWASP API Security best practices
• CI/CD: GitHub Actions, Argo CD, Docker
• Monitoring & Logging: Prometheus, Grafana, ELK Stack (Elasticsearch, Logstash, Kibana)

• Modular core. We rebuilt the survey platform as three services — orchestration, response intake, and insights, running on AWS ECS Fargate with Apache Kafka for event streaming. This lets us scale ingestion, analytics, and delivery independently without disrupting user experience.
• Adaptive flows. We implemented a PostgreSQL rules engine to personalise question paths in real time using HRIS data. We used Redis to cache templates, keeping response times under one second.
• AI-driven analysis. We processed free-text answers with spaCy and Hugging Face models for sentiment, topic, and intent, then summarised them through OpenAI GPT-4.5 Enterprise with citations and bias checks. We stored embeddings in Elasticsearch to enable instant semantic search across historical survey data.
• Live insights. We built GraphQL subscriptions to push real-time sentiment changes, anomalies, and completion rates directly to React + D3.js dashboards. We integrated Apache Superset for deeper drill-downs.
• Secure and compliant. We tokenised PII at intake, encrypted it with AWS KMS, and stored it regionally. We enforced RBAC for access to raw data and summaries, and anonymised all AI calls for audit readiness.
• Reliable delivery. We set up GitHub Actions and Argo CD for blue-green deployments with zero downtime. We managed schema changes via versioned migrations and fallbacks to ensure stability.
• Full observability. We utilized Prometheus, Grafana, and ELK to monitor latency, inference times, and cost metrics, ensuring performance and budget remained fully under control.
• Action-oriented UX. We designed explainable summaries, trend comparisons, and recommended actions with confidence scores so that managers could make immediate, informed decisions.

BUSINESS OUTCOMES

• Participation Lift. Engagement rates rose from 52% to 83% through real-time, role-aware question flows.
• Decision Speed. HR gained same-day visibility, with complete datasets processed in 90 minutes post-close.
• Retention Stability. Early sentiment detection improved retention by 12% in at-risk teams.
• Manager Enablement. Auto-generated action packs delivered in Slack/Teams equipped managers to act immediately.
• Trust Increase. Transparent bias scoring and anonymisation elevated employee confidence in feedback integrity.

TECHNICAL OUTCOMES

• High-Scale Processing. Multi-tenant engine sustained 1.2M+ events monthly without SLA breaches.
• Real-Time Pipeline. Kafka + ksqlDB ingestion reduced submission backlogs by 78%.
• AI Efficiency. Fine-tuned GPT-4o with RAG context delivered 200+ summaries in parallel, averaging 2.4 s per run.
• Zero-Downtime Delivery. Blue-green Kubernetes deployments kept uptime at 99.985%.
• Data Security. Field-level encryption with per-tenant AWS KMS keys ensured GDPR/CCPA compliance.
• Model Quality Control. LangSmith + Prometheus continuously monitored AI accuracy, coverage, and bias.