Enterprise Knowledge Base System

Overview

Architecting an organization-wide intelligent knowledge base designed to serve thousands of professionals across a large global engineering consultancy. The system transforms decades of accumulated engineering expertise, internal documents, and project databases into a queryable, self-improving institutional memory. Engineers and professionals across disciplines can surface relevant knowledge in seconds rather than searching through siloed repositories or relying on tribal knowledge.

The project is deeply research-driven at every layer: from selecting the right on-premise language model to serve at scale, to evaluating the optimal vector database architecture, to designing a feedback loop that allows the knowledge base to grow smarter with every interaction.

My Role

Lead AI Engineer and Researcher — Conducting intensive research across the full AI stack to make the right architectural decisions before committing to production infrastructure. Responsible for model evaluation, retrieval architecture, data ingestion pipeline, and the long-term knowledge accumulation strategy.

Research and Technical Architecture

On-Premise LLM Serving

• Model Evaluation: Intensively testing Qwen and Gemma families on on-premise GPU infrastructure to identify the best performing model for engineering domain queries, with a planned fine-tuning phase on the organization's proprietary document corpus
• VLM Research: Evaluating Vision Language Models (VLMs) for intelligent parsing of engineering documents, technical drawings, and scanned reports — a critical capability given the diversity and complexity of the document formats involved
• Serving Infrastructure: vLLM as the high-throughput inference engine, enabling efficient concurrent serving across the organization while keeping sensitive data fully on-premise

Retrieval Architecture

• Vector Database Selection: Conducting rigorous benchmarking of leading vector stores (including Weaviate and Qdrant) to determine the best fit for hybrid search requirements, query volumes, and long-term scalability needs
• Document Store: MongoDB as the primary document store for metadata management, raw content persistence, and flexible schema handling across heterogeneous document types
• Knowledge Graphs: Building a graph-based layer to capture relationships between projects, disciplines, standards, and domain concepts, enabling traversal-based retrieval that goes beyond pure vector similarity
• RAG Orchestration: Evaluating LangChain, LlamaIndex, and a fully custom pipeline against specific performance, flexibility, and long-term maintenance requirements

Accumulative Wiki and Self-Improvement

• Recursive Language Models (RLM): Researching RLM-based approaches for hierarchical semantic parsing of complex engineering documents — enabling the system to build tree-structured representations of nested technical concepts, clauses, and multi-part specifications that flat sequence models cannot fully capture
• Knowledge Graph Evolution: Designing the graph schema to grow organically as new projects, standards, and documents are ingested, making institutional knowledge an ever-expanding connected graph rather than a static index
• Accumulative Architecture: Every query, correction, and expert validation feeds back into the system, gradually building a living knowledge base that reflects the organization's evolving expertise across multiple engineering disciplines

Data Sources and Infrastructure

• Multi-Source Ingestion: Unified pipeline ingesting from SharePoint repositories, internal relational databases, engineering document archives, project management systems, and technical standards libraries
• Hybrid Deployment: Core LLM inference and sensitive data processing on on-premise servers; scalable orchestration and auxiliary services on Azure
• Document Coverage: Engineering reports, feasibility studies, cost analyses, technical specifications, and project records spanning multiple decades and disciplines

Scale and Scope

• Organizational reach: Designed to serve thousands of engineers, architects, planners, and project managers across global offices
• Domain breadth: Infrastructure, transportation, airports, metro systems, water treatment, buildings, urban planning, and energy sectors
• Knowledge depth: Decades of project records from landmark global projects spanning multiple continents
• Security-first: Sensitive engineering IP stays on-premise; architecture designed around data residency and access control requirements

Technical Skills Demonstrated

Research Approach

Building an enterprise knowledge base at this scale demands rigorous research before every architectural decision. The wrong vector database choice at production scale is expensive to reverse. A poorly chosen LLM burns GPU cycles without delivering accuracy. An orchestration framework that fits a demo but breaks under organizational data diversity becomes a liability.

The methodology is systematic benchmarking: each candidate technology is evaluated against realistic samples of the organization's document corpus, query patterns, and performance requirements. Model fine-tuning research focuses on domain adaptation for engineering terminology, multi-language technical content, and document-heavy retrieval tasks including parsed drawings and scanned reports via VLMs.

The accumulative wiki vision is the most ambitious part: designing a system where every approved expert answer becomes a node in the knowledge graph, and RLM-based hierarchical parsing progressively improves the semantic precision of retrieval — capturing nested technical relationships without requiring manual curation at scale. The goal is a knowledge base that becomes measurably smarter the more it is used.