Cost Engineering Multi-Agent System

Professional Work | 2025 - Present

Note: This work is confidential under NDA. Details presented focus on general engineering skills and publicly available technologies without disclosing proprietary information or client identity.

Overview

Built and deployed a production 10+ node multi-agent system using LangGraph, serving the cost control department of a large global engineering consultancy. The system allows cost engineers to query decades of historical project cost data through natural language, analyze Bills of Quantities (BoQs), and generate reliable, SQL-grounded cost estimates for major infrastructure projects.

The client operates at the scale of some of the world's most complex infrastructure programs. Their cost control team manages data across a vast portfolio of completed and active projects globally. This agent brings that accumulated knowledge to engineers' fingertips, replacing manual database queries and spreadsheet hunting with a reliable, conversational interface backed by verifiable SQL results.

My Role

Lead AI Engineer — Designed the full multi-agent graph topology, implemented all reasoning nodes and tool integrations, established the SQL query validation pipeline, and deployed the system to serve the cost control department.

Agent Architecture

The system is built as a directed graph of 10+ specialized LangGraph nodes, each responsible for a distinct reasoning or execution step. The topology ensures that every final answer is grounded in verifiable data, not model hallucination.

Core Agent Nodes

Model Strategy

Key Capabilities

Cost Intelligence

Trustable Numbers

Scale and Context

Technical Skills Demonstrated

LangGraph Multi-Agent Systems Text-to-SQL SQL API Integration Large Language Models LLM Fine-tuning BoQ Analysis Cost Engineering Python FastAPI Agent Orchestration Prompt Engineering Data Validation Structured Output Parsing Azure

Engineering Learnings

The core challenge in a cost engineering agent is not accuracy in isolation, it is trust. A cost engineer staking a project budget on an AI-generated figure needs to know exactly where that number came from. Designing the answer verification and source attribution nodes was as important as the reasoning capability itself.

Managing a 10+ node graph requires deliberate topology design. Early cycles revealed that routing quality at the entry node had outsized impact on the entire pipeline. An incorrectly routed query that reaches the wrong sub-agent produces plausible-looking but wrong answers. The query router was the node that received the most iterative refinement.

The transition from a mixture of models toward a fine-tuned specialist model is driven by the same principle: general-purpose LLMs work, but they occasionally struggle with domain-specific BoQ terminology, project type classification, and the structured SQL patterns that the cost database requires. A fine-tuned model should close that gap significantly.

Back to Projects