🤖 HR Policy RAG Chatbot

Conversational RAG chatbot using Azure OpenAI GPT-4o mini, LangChain, and Azure AI Search — ingesting 500+ policy documents via Azure Databricks (PySpark + spaCy), hybrid BM25 + dense retrieval, MMR re-ranking, and multi-turn memory. Reduced HR support ticket volume by 38%. Deployed on AKS with Azure DevOps CI/CD; RAGAS evaluation gate on every push; Azure Redis Cache cutting API costs by 40%.

⚡ Payroll Anomaly Explanation Engine

LLM explanation layer on PayCom's rules-based anomaly detection system — enriching flags with 6-period payroll history and HR event context from Azure SQL + Cosmos DB, generating structured plain-English investigation summaries via GPT-4o JSON mode. Deployed as Azure Functions triggered by Azure Service Bus. Improved admin action rate on anomalies by 71%.

💬 Employee Self-Service AI Assistant

Two-layer conversational assistant: Layer 1 fetches live personal payroll data from Azure SQL scoped to the authenticated employee; Layer 2 retrieves policy context from Azure AI Search. LangChain router classifies query type in real-time; Redis session memory for multi-turn context; WebSocket token streaming via React.js. Handles 2,000+ daily queries with 84% self-resolution rate — reducing HR workload by 45%.

📄 Document Intelligence Pipeline

Automated extraction pipeline for W-2s, I-9s, offer letters, and contracts — Azure Document Intelligence + OpenCV pre-processing (OCR accuracy 67% → 89%), GPT-4o function calling with Pydantic schemas, Map-Reduce for multi-page contracts, 3-tier validation layer. Reduced document processing time by 60% and data entry errors by 72%.

🔍 Agentic Payroll Audit Agent (LangGraph POC)

Autonomous audit agent using LangGraph — 6-node state graph: AnomalyAnalyser → DataFetcher (5 tools) → EvidenceEvaluator → RootCauseDeterminer → RecommendationGenerator → AuditReportWriter. Conditional edge loops back to DataFetcher if evidence insufficient (max 3 iterations).

Shipped 5 production AI features on DarwinBox's enterprise HCM platform, serving 30+ global clients including Swiggy (10,000+ employees) — owning backend data pipelines, prompt engineering, and LLM integration across the Performance Management System.

✅ Feedback Summarisation

Backend pipeline for LLM-generated 360-degree feedback summaries — processing reviewer comments across reviewer_id/reviewee_id mappings with TF-IDF scoring to fit token budgets, enabling managers to consume multi-stakeholder feedback in seconds instead of hours.

✅ Auto Goal Generation

Context-assembly pipeline aggregating employee role, historical performance, and team OKRs from MongoDB into structured LLM prompts. Nightly batch cron pre-computes context objects; Pydantic validation enforces SMART goal structure with JSON mode output.

✅ Pre-populate Review Comments

Prompt construction pipeline retrieving past performance records, goal scores, and peer feedback via MongoDB aggregation pipelines. UX framing change (first-person → observation style) reduced manager rejection rate from 41% to 12%.

✅ MSF Summary

Multi-stakeholder feedback pipeline with proportional stakeholder sampling and TF-IDF scoring when context exceeded token limits — delivering narrative LLM summaries across enterprise review cycles.

✅ Skill Generation

AI-driven skill inference pipeline extracting signals from goal data, feedback keywords, and competency ratings via MongoDB aggregation. Taxonomy compliance validation ensures all inferred skills map to a 450-skill company taxonomy. Adopted by HR teams for talent gap analysis.

🔧 Infrastructure: WiredTiger compaction reduced MongoDB from 150 GB → 500 MB (99.7% reduction), query times from minutes → under 5 seconds. Cron refactoring from single-record to batch processing cut runtime from never-completing → under 10 minutes per client.