Client: Integrum Ops — builder of TMG Project Center (TMGPC), a new-to-market commercial construction project-management platform for general contractors and major-trade subcontractors.
Engagement: End-to-end AI/ML capability — Jacobian designed and built the TMGPC AI Agent on AWS.

The problem

Commercial construction project management is a wide, heterogeneous surface. A single job touches submittals, RFIs, drawings, daily reports, schedules, punch lists, and a constellation of microservices that each own a slice of the truth. A monolithic AI prompt can’t reason coherently across all of it — and, more importantly, it can’t safely act on a live platform without human oversight and bounded authority.

Integrum Ops needed an agentic assistant that could answer complex cross-domain questions, take actions on the platform, and do both in a way that was auditable, governable, and ready for production — not a weekend prototype.

What Jacobian built

Jacobian engineered a production multi-agent system purpose-built for TMGPC’s architecture.

At the core is a supervisory agent backed by a capability registry. When a user request arrives, the supervisor reads the registry and delegates to one or more domain-expert agents, each scoped to a slice of the platform — submittals, RFIs, schedule, documents, and so on. Each expert agent is equipped with typed LangChain tools that wrap the corresponding TMGPC microservice APIs, so it can only touch what it’s authorised to touch. Orchestration is handled by LangGraph, exposed via a FastAPI service deployed in Docker on EKS spot node groups.

All foundation-model calls are routed through AWS Bedrock — no direct model API calls anywhere in the stack. Bedrock provides a single governance and cost-control plane and keeps model choice flexible without rewriting application logic.

Documents get their own dedicated pipeline. Upload events arrive on Kafka, triggering async AWS Textract extraction. Extracted content flows into a metadata and keyword index in RDS PostgreSQL, while Amazon Titan (via Bedrock) generates embeddings stored in ChromaDB — a full agentic RAG pipeline that lets the agents reason over project documents at query time.

When a tool call returns ambiguous results — multiple matching submittals, say, or several relevant drawing revisions — the system uses LangGraph interrupt/resume to surface the choice to a human operator before proceeding. Checkpoint state is persisted in PostgreSQL, so the conversation resumes exactly where it paused regardless of which container picks it up.

The outcome

TMGPC launched with a trustworthy, governable agentic assistant embedded in the platform: expert agents with bounded authority, an auditable Bedrock LLM plane, human-in-the-loop resolution for ambiguous actions, and spot-instance resilience baked into the deployment.

“Jacobian built this with us as a real engineering partner — multi-agent orchestration, Bedrock governance, the human-in-the-loop flow. They treated it as a production system from the first commit, not a proof of concept.”
— Chris Worth, Co-Founder & Chief Architect, Integrum Ops

It’s how Jacobian builds production agentic AI: bounded agents, auditable infrastructure, and human oversight where the stakes demand it.

Read the full engineering story → The complete Integrum Ops case study on TrustEdge