AI-Based Closed Loop Engineering for Map Compiler

Project snapshot

In scope of this internal R&D initiative, Intellias designed and validated an AI loop engineering architecture for map compiler, combining context engineering, task planning, implementation, review, and validation agents to generate and refine compiler code. The goal was to test how far AI-driven code generation could be pushed for the client’s map compilation pipeline: from assisted development, where a human owns every step, to a closed-loop system, without compromising the quality and reliability that map compiler code demands. This loop engineering software development approach resulted in a working proof of concept delivered in three weeks.

Business challenge

• High costs and long timelines – Map compilation solutions require significant engineering effort whenever a new input or output data format is introduced, driving up cost and delaying delivery.
• Proprietary constraints – The client’s existing compiler codebase relies on proprietary frameworks and domain-specific libraries, which limits straightforward reuse and automation.
• Feasibility question – The client needed to understand how far AI-driven code generation could realistically be pushed for this class of problem — from assisted development to autonomous closed-loop generation — without compromising quality and reliability.

Solution

Domain-aware RAG pipeline

Before generating any code, Intellias built a domain-aware Retrieval-Augmented Generation (RAG) pipeline that preprocesses:

• the client’s input and output format specifications
• the existing compiler codebase
• domain-specific libraries and APIs

Git history

The pipeline uses Docling, Context7 MCP, Serena MCP, ChromaDB, and Memgraph to prepare and serve this context to the agents, giving them deep contextual understanding of the target codebase.

Agentic orchestrated setup

The architecture coordinates four loop engineering AI agents, built with Claude Code subagents, across the software development lifecycle. This Claude Code loop engineering approach connects planning, implementation, code review, and validation within a shared workflow:

• Work Planner Agent – built on Claude Code subagents and OpenSpec, this agent turns requirements into a task list and implementation plan, drawing on context prepared by the RAG pipeline (Context7 MCP, Serena MCP).
• Implementation Agent – built on Claude Code subagents, this agent writes the code against the plan.
• Code Review Agent – built on Claude Code subagents, this agent reviews the pull requests the Implementation Agent generates; pull requests move between the two agents in an iterative loop.
• Validation Agent – intended to close the loop by evaluating implementation progress and returning validation results to the Implementation Agent. This loop engineering agent was scoped for a later phase and was out of scope for this PoC.

Together, these components demonstrate how agentic loop engineering, implemented through Claude Code subagents, can support a structured and repeatable approach to complex software development.

Business outcome

Rapid PoC delivery

Intellias delivered a functional PoC in three weeks, validating the architecture by compiling a real map, including POI and road data.

Technical milestone

The PoC generated and evaluated a real map compiler module, producing a compiled map in NDS format, code pull requests, and test results to quantify efficiency gains.

Architectural reusability

The closed-loop code generation architecture is proven and reusable: Intellias designed it to be applied across other map compilation components and future data formats, accelerating subsequent implementation of Claude loop engineering practices.

Efficiency gains

The AI loop engineering approach is estimated to reduce development effort and delivery time by 40% when onboarding new geospatial data inputs into the client’s map compilation pipelines.

Technologies: Java, GeoTools, Shapefile, Claude Code, MCP, Context7 MCP, Serena MCP, Docling, ChromaDB, Memgraph, OpenSpec, NDS