A milestone post — CSI-AI v0.7 beta is end-to-end. From a single chat window, the system goes from "describe a building" all the way to "here are the analysis results, structured for an engineer to act on." Every phase of the original five-phase plan is operational. SAP 2000 and ETABS both supported with the same chat experience.
The full pipeline, working
What end-to-end means in practice:
- Model creation — geometry, joints, frames, areas, all from prompt
- Material assignment — steel grades, concrete strengths, density and modulus consistent with the named grade
- Load application — patterns, combinations, area-uniform loads, frame point loads, all addressable via natural language
- Analysis execution — kicked off from chat, progress tracked, errors surfaced back into the conversation
- Results extraction — base reactions, member forces, displacements pulled back into context, summarized for the engineer
The thing that took a while to get right was the last step. It is one thing for the LLM to read a CSV of analysis results; it is another to surface those results in a way an engineer would actually read in a project meeting. The post-analysis summarization had to be wired carefully — too verbose and it is unreadable, too terse and it loses critical context.
Smart automation pieces
Underneath the pipeline, four automation surfaces matured at the same time:
- Intelligent material creation and assignment — the LLM picks named grades from a curated library and falls back to custom definitions only when the engineer’s description requires it
- Comprehensive loading system with combinations — load patterns, factored combinations, ASCE 7 envelope handling
- Automated analysis execution — solver invocation, status polling, error recovery if the analysis fails to converge
- Real-time results interpretation — outputs piped back into the conversation, summarized with units and sign conventions called out explicitly
Both SAP 2000 and ETABS, same chat
The earliest prototype was SAP-only. v0.7 is the first version where the same prompt, the same chat experience, drives either tool. Most of that came down to the fact that CSI Labs ships a remarkably consistent OAPI across the family — the underlying tool catalog only had to fork in a handful of places. That portability is what makes the whole multi-product play viable.
Honest list of what still needs work
- Modifying existing structures still struggles in places — the same rough edge from earlier phases. Working on it.
- Performance is uneven across operation types. Some calls are sub-second, others stall at random and need investigation.
- No parallel processing yet — every tool call serializes. There is a real speedup waiting in batched, parallel OAPI calls for independent objects.
- Currently developing a custom MCP client interface so the prototype is not tied to Claude Desktop forever.
- Compatibility expansion beyond the CSI family — Robot, Revit, eventually STAAD — is the next horizontal axis to push on.
What is coming next
Two projects in flight:
- Civil Core Brain — a reasoning layer on top of the chat that handles longer-horizon planning. The LLM decomposes "design a 12-story office" into a sequence of phases the prototype can execute, and re-plans on the fly when something does not converge.
- Civil Core Agent — a more autonomous mode where the chat operates closer to a structural-design assistant than a tool-calling chatbot. Less "you say, it does"; more "you state the goal, it asks the right questions, then does."
Both names are working titles — they will probably get rebranded by the time anything ships. (They eventually became ConGro AI.) The point of mentioning them now is to flag that the next post will be about the reasoning architecture, not just the tool catalog.
End-user release approaching
This is the first version where I can imagine handing it to a non-me-shaped engineer and watching them get value out of it inside a single hour. That is the threshold I have been chasing. End-user release is no longer a theoretical milestone — it is the next thing on the roadmap.
If you are an engineer who would want to be on the early-access list, the easiest thing is to follow along. There will be a sign-up surface soon. In the meantime, feedback on what is shown above is genuinely useful — what would you want this thing to be able to do that it cannot yet?
Thanks
Computers and Structures, Inc. — thank you for the encouragement. The tooling and the OAPI design are what make all of this possible. Looking forward to what comes next.