Lishan – LLM, RAG, LangChain

Metadata-driven enterprise ELT platform for a major German energy company, integrating dozens of source systems into a centralized Azure data platform following a medallion architecture, with Palantir Foundry and Power BI as downstream consumers.

• Operated and extended a metadata-driven enterprise ELT framework built on Azure Data Factory, Azure Databricks, ADLS Gen2, Unity Catalog, Delta Lake and Azure SQL Database, following a medallion architecture (Landing/Raw/Cleaned/Derived) with Palantir Foundry and Power BI as downstream consumer layers;
• Developed and maintained the Azure SQL-based metadata and configuration layer driving all ingestion and transformation pipelines: T-SQL schema design, views, computed columns and state-driven orchestration logic, working inside an existing model owned by the platform's data architect;
• Onboarded new source systems end-to-end, including the platform's first SAP HANA HT2 (SLT-based) connection: config-table design in Azure SQL, Key Vault credential setup, network approvals and schedule integration;
• Conducted a cross-system root-cause analysis of source-delivery and scheduling chains across 162 tables by correlating Blob Storage landing timestamps, Delta Lake transaction history and ADF trigger configurations, producing a reusable Databricks diagnostic notebook and concrete remediation proposals;
• Built and extended config-driven ingestion and cleaning pipelines, including KPI push pipelines exporting curated datasets to Palantir Foundry with append/update merge logic and data-quality validation (primary-key uniqueness, null checks, delta hash monitoring);
• Decommissioned legacy ADF pipelines and Databricks repositories after platform migration, including dependency analysis of downstream consumers, trigger shutdown via CLI and handover documentation;
• Coordinated framework changes across source-system owners, framework admins and data-product owners, distinguishing quick wins from source-side changes requiring multi-party alignment.

Production AI/ML pricing and inventory intelligence platform for a US heavy-equipment dealership group, combining predictive models, data from multiple operational and external sources, and an AI-powered advisory agent to support inventory and sales decisions.

• Productionized a machine-learning scoring model on Azure, implementing automated batch scoring, scheduled retraining, forward validation, MLflow experiment tracking, model registration, and controlled promotion through Azure Machine Learning;
• Developed a two-stage CatBoost solution for sell-speed risk, combining a hedonic market-value regressor with a multiclass days-to-sell classifier, isotonic probability calibration, and time-based validation across 100,000+ equipment records;
• Engineered data pipelines integrating 11 operational and external data sources across Azure Synapse serverless SQL, Cosmos DB, marketplace feeds, Google Analytics 4, and Salesforce into a unified analytical layer for ML scoring and downstream applications;
• Built Python/FastAPI services exposing ML predictions, inventory analytics, pricing intelligence, and AI-generated insights to downstream applications and business users;
• Architected a cloud-native application stack on Azure Container Apps, combining a FastAPI backend, React dashboard, Streamlit trade-in valuation app, Azure Functions, and Azure Front Door with WAF;
• Built a LangChain/Azure OpenAI advisory agent generating pre-call briefs for salespeople from service history, quotes, marketplace comparables, inventory context, and ML predictions;
• Established CI/CD, automated testing, monitoring, Redis caching, cache warming, and Locust stress testing to support reliable and scalable production operations;
• Provisioned production, demo, and development environments through modular Bicep IaC, including private networking, NAT/static egress, Key Vault, custom domains, TLS, and passwordless Managed Identity authentication.

AI-powered ticket intelligence and knowledge platform for a pharmacy IT company, combining large-scale helpdesk and work-item data with automated duplicate detection and a Microsoft 365 Copilot grounded in enterprise knowledge sources.

• Built and operated production Databricks data pipelines using Unity Catalog, Delta Lake, and a Bronze/Silver/Gold architecture, processing 100,000 helpdesk tickets, 1.5 million interactions, and 7,300 Azure DevOps work items;
• Implemented automated ingestion and transformation pipelines integrating Databricks, SharePoint, Azure AI Search, and downstream services, with incremental synchronization, ETag-based change detection, and idempotent processing;
• Engineered a Python/FastAPI AI ticket-matching microservice to identify whether incoming helpdesk tickets duplicate existing Azure DevOps work items;
• Orchestrated a four-stage matching pipeline combining Databricks SQL input resolution, GPT-based query normalization, Azure AI Search hybrid retrieval (vector, keyword, and semantic reranking), and LLM-as-judge evaluation with confidence scores and explanations;
• Designed a Microsoft 365 Copilot knowledge architecture grounded in Azure AI Search, providing cited answers with deep links to SharePoint source documents;
• Established CI/CD, automated testing, tracing, quality and coverage gates, and API authentication to support maintainability, observability, and production reliability.

Enterprise RAG knowledge assistant for a public-sector client, providing context-aware, cited answers across large document collections through hybrid retrieval and multi-agent orchestration on Google Cloud.

• Built a high-performance embedding pipeline with intelligent chunking, automated index synchronization, and high-watermark change detection, using Vertex AI to efficiently process incremental document updates;
• Engineered a production-grade RAG system using Vertex AI Search for hybrid retrieval (vector, keyword, and semantic reranking) and Vertex AI Vector Search for low-latency ANN retrieval, with Google’s Agent Development Kit (ADK) for multi-agent orchestration and intent extraction;
• Developed a multi-container microservices platform on Cloud Run with Pub/Sub-driven autoscaling, FastAPI backend services, and Streamlit interfaces, provisioned through Terraform for reproducible deployments.

AI-powered lead-generation platform for a consultancy, combining enterprise knowledge with a RAG-based chat experience to answer domain-specific queries, identify high-intent prospects, and feed qualified leads into the sales pipeline.

• Architected an end-to-end document ingestion pipeline using LangChain-based Map-Reduce LLM summarization and automated classification agents, triggered by Azure Blob Storage to process enterprise documentation;
• Built an LLM-based classification agent to categorize documents into domain-specific indexes (Sales, Knowledge, Facts) with confidence scoring, using MongoDB-backed change detection to maintain data integrity;
• Developed a high-performance embedding pipeline with intelligent chunking, automated index synchronization, and high-watermark change detection for incremental document updates;
• Engineered a production-grade RAG system using Azure AI Search for hybrid retrieval (vector, keyword, and semantic) and Azure OpenAI embeddings, with LangGraph for intent extraction and context-aware responses;
• Developed a multi-container microservices platform on Azure Container Apps with queue-driven autoscaling, FastAPI backend services, and Streamlit interfaces, provisioned through Bicep for reproducible deployments;
• Implemented a GDPR-compliant PII masking pipeline using Azure Text Analytics to redact names, emails, contact details, and other sensitive information before database persistence;
• Integrated lead-capture workflows into the AI chat experience, correlating session data with MongoDB to identify high-intent prospects and feed contact details into the sales pipeline.

Enterprise AI support chatbot for a mass-media company, providing secure, context-aware answers grounded in internal documentation and ticketing data through RAG and vector search.

• Engineered an enterprise AI support chatbot using Azure OpenAI, PromptFlow, and Azure AI Search to provide contextual answers grounded in internal documentation and ticketing data;
• Developed automated ingestion pipelines for Confluence and ServiceNow, extracting, processing, and indexing enterprise content for AI-powered knowledge retrieval;
• Implemented secure vector and semantic search using Azure AI Search and custom embeddings, designing optimized index schemas and configurable data sources, indexes, and indexers with high-watermark synchronization;
• Built and optimized PromptFlow pipelines for intent extraction and contextual response generation, with a modular architecture and production monitoring;
• Developed a PII detection and masking pipeline using Azure Language services, preventing sensitive information from being persisted or exposed to the chatbot and search index;
• Built and deployed FastAPI backend services and a Streamlit chat interface on Azure Container Apps using Docker, supporting both custom frontend and interactive chat experiences;
• Established GitLab CI/CD to automate testing, integration, and deployment workflows.

Cross-border financial-reporting automation platform for an energy-sector enterprise, digitizing GAAP reporting across 250+ legal entities through specialized AI agents, automated quality controls, and enterprise data infrastructure.

• Conceptualized the end-to-end solution architecture for a cross-border financial automation platform digitizing GAAP reporting across 250+ legal entities;
• Designed a high-availability hybrid architecture integrating Databricks (Unity Catalog) with Azure OpenAI, using Managed Identities and zero-trust principles to meet GDPR and energy-sector security requirements;
• Architected a multi-agent orchestration layer using LangGraph, defining handoffs, state management, and exception handling across Financials, Regulatory Tables, and Narrative Generation agents;
• Designed a six-gate Quality Control framework incorporating balance-sheet reconciliation, cross-agent validation, and LLM-based quality scoring to support audit-ready outputs;
• Standardized cross-platform integration through a FastAPI service layer, enabling low-latency communication between Databricks Model Serving and the Azure orchestration layer;
• Developed the five-year strategic roadmap and TCO model supporting the platform investment decision;
• Defined the operational governance model (RACI) across Data Engineering, Platform Operations, and Finance to support a 99.9% uptime SLA during critical month-end closing windows;
• Produced architectural blueprints, including sequence diagrams, data-flow models, and authentication schemas, supporting both technical implementation and executive-level review.