Local-only runtime
Principle: no external API, telemetry, or CDN — run on the workstation and internal network so AI execution keeps data on site.
Product
nanochat
A local Engineering Agent Runtime for the CAD workstation.
Under the premise that engineering data does not leave the site, we are designing a runtime that joins a local 3D viewer, OID-backed GraphRAG, and a local SLM. It is not a cloud chatbot replacement.
Design stage. Product code is not shipped yet; the following describes target architecture and principles in the future tense. No GA or measured savings claims.
Problem
As AI use grows, so does the On-prem GPU infrastructure burden.
EPCs and Owner Operators commonly operate design data On-prem. As AI adoption expands, so does the burden of centralized GPU infrastructure and large-LLM operation. nanochat is being designed to use Engineering Context and GraphRAG to narrow the work required from the model and enable local AI Agent execution on existing CAD workstations.
What it will do
Join local Viewer + GraphRAG + SLM into one agent.
nanochat targets a fully local 3D viewer agent on EPC and shipbuilding CAD workstations (roughly 8–16GB NVIDIA class). Users will ask through a side panel or in-viewer chat; the agent will answer from a Smart3D-derived spatial graph and drive the viewer.
- Viewer: CesiumJS + 3D Tiles + local bridge
- Knowledge: Smart3D extract → GraphRAG (SQLite); factual claims land on OIDs
- Inference: local SLM — no external API, telemetry, or CDN
Target I/O flow
Engineering Context → OID-backed response
One round-trip: context-backed local SLM calls only whitelisted tools, drives the viewer, and answers with OID-backed factual claims.
- 01
Engineering Context
Process and spatial graph ground truth
- 02
GraphRAG
OID-backed retrieval
- 03
Local SLM
Intent · tool choice
- 04
Whitelisted tools
No arbitrary SQL or code generation
- 05
Viewer action
Highlight · isolate · camera
- 06
OID-backed response
Verifiable factual claims
Capabilities in design
Pillars defined toward Alpha
OID-backed GraphRAG
Factual claims land on engineering object OIDs. The SLM will not generate SQL or arbitrary code; it will call whitelisted tools only.
Two surfaces, one protocol
Browser side panel and in-viewer chat are designed to share the same conversation engine and tools — no feature that works on only one surface.
Design-review assistance
Target scenarios include lookup, visibility, spatial neighborhood, and hierarchy summary — locally. Not autonomous approval or a cloud chatbot substitute.
Target workflow
Paths Alpha aims to validate
- 01
Place local server, viewer, and graph DB on the CAD workstation.
- 02
Ask in Korean through the side panel or in-viewer chat.
- 03
The agent queries the graph via whitelisted tools and cites OIDs.
- 04
The viewer highlights, isolates, and fits the same targets.
- 05
Validate the same scenarios with the network blocked.
Boundaries and authority
What design-stage copy will not claim
Because product code is not shipped, we do not treat performance, savings, or fully offline guarantees as facts. AI has no approval authority.
| Layer | Role |
|---|---|
| Engineering Context / Graph | Ground truth (OID) — nanochat does not redefine it |
| Local SLM | Intent, tool choice, answer generation |
| Viewer | Visualization and manipulation surface |
| Engineer | Final review, disposition, and approval authority |
- No GA or production-ready claims (Alpha planned date only).
- No measured time or cost savings figures.
- Not positioned as a cloud chatbot or generic LLM portal replacement.
- No promise of autonomous approval or absolute offline guarantee.