CAE Copilot — Engineering analysis operating layer¶

Data-first knowledge · traceable agents · human-orchestrated CAE workflows

Enterprise AI adoption strategy for engineering organizations — govern data first, deploy agents with citations and deterministic tools, keep the engineer as orchestrator. The shipped portfolio demo is the ANSYS Rocky (DEM) adapter — cited answers from v26r1 manuals, not a generic chatbot.

Start here

Problem statement and positioning — platform vision, five engineering problem areas, Rocky adapter scope, 3-stage maturity model. Read this before architecture details.

Stage	Focus	Status
1 — Data Foundation	Rocky PDF ingest, vector RAG + metadata filters	Done (MVP corpus)
2 — Trusted Agent Layer	LangGraph, tools, HITL, eval	Done
3 — Solver Orchestration	Async CAE jobs	Deferred

Details: status · roadmap.

Live endpoints¶

Surface	URL
UI	cae-copilot.dmitryivanov.dev
API (Swagger)	api.cae-copilot.dmitryivanov.dev/docs
Docs	docs.cae-copilot.dmitryivanov.dev

Usage economics¶

The agent layer is model-agnostic — orchestration, citations, and tools are the product; the LLM is a replaceable inference layer. The public demo runs on OpenRouter with openai/gpt-4o-mini (local dev may use a free-tier model; see quickstart).

	Production (OpenRouter)
Provider	`CAEC_LLM_PROVIDER=openrouter`
Model	`CAEC_LLM_MODEL=openai/gpt-4o-mini`
Input	$0.15 / 1M tokens
Output	$0.60 / 1M tokens
Typical LLM calls	2 (RAG path) or 3 (calculation path)

Per request, the graph calls the LLM for intent on every path, then report synthesis (RAG/setup/scripting) or tool selection + report (calculation; up to three tool retries on validation errors). Input tokens grow with retrieved chunks (CAEC_RAG_TOP_K, default 5); output is capped by CAEC_LLM_MAX_TOKENS (default 2048). Embeddings run locally (sentence-transformers on CPU/GPU) — only chat completion is metered.

Illustrative costs (not a benchmark — verify with your own traces):

RAG answer (~2.5k input + ~600 output tokens): ~$0.001 per request
Calculation path (~4k input + ~1k output tokens): ~$0.002–0.003 per request

On-prem option: point CAEC_LLM_BASE_URL at an OpenAI-compatible endpoint on internal GPU (vLLM, LM Studio, corporate Qwen/YandexGPT-class gateway) — no graph changes required. Qdrant and embeddings already deploy on-cluster (deployment). Details: LLM providers.

No API key

Without CAEC_LLM_API_KEY, the graph uses keyword intent and regex tools — zero LLM cost, reduced answer quality. Fine for CI and offline development.

Prices change — verify on OpenRouter.

Quick start¶

mamba env create -f environment.yml
mamba activate cae-copilot
pdm install
pdm run test
pdm run api

See quickstart.md for Qdrant, ingestion, and deployment notes.

Portfolio: overview · problem statement · pitch.

Traceable flow (Stage 2)¶

Core loop (one branch per request):

User query → Intent → Router → [RAG | Tools | (direct)] → Report

Demo paths — four scripted scenarios (one branch per request):

flowchart TD
    Q[Question] --> I[Intent]
    I -->|knowledge setup scripting| R[RAG retrieval]
    R --> C[Citation in report]
    C --> Rep[Report]
    I -->|calculation| T[Calculation tool]
    T --> Rep
    I -->|unknown| Rep

See demo walkthrough for the scripted demos.

Architecture: overview · Data-First.