Skip to content

Pitch

Short summary for recruiters and interview openers. Full narrative: problem-statement.md. Wording cheat sheet: positioning.md. Disclaimer: disclaimer.md.


Product vision

CAE Copilot — Engineering analysis operating layer

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

Not a chatbot. An enterprise adoption strategy for engineering AI: govern data first, route agents with citations and deterministic tools, keep the engineer as orchestrator. The live demo is the Rocky (DEM) adapter — first pluggable product on the platform.


For technical interviewers (30 seconds)

Fourteen years in CAE/HPC taught me that engineering answers need traceability and validation. CAE Copilot is an engineering analysis operating layer — Data-First routing, LangGraph agents with citations, Pydantic-validated DEM tools, and a human orchestrator for high-risk outputs. Stage 1 (Rocky PDF corpus) and Stage 2 (agent layer) are shipped in prod; Stage 3 solver orchestration is designed but deferred. Same reliability mindset as my production ML work: eval, guardrails, and scope honesty.


For CTO / Chief Engineer

Engineering organizations often fail at AI adoption in predictable ways: agents deployed on unstructured document chaos, GUI macros that automate bad AS-IS workflows, black-box answers without citations, and lock-in to a single cloud LLM or CAE vendor.

CAE Copilot is a 3-stage maturity roadmap, not a demo chatbot:

  1. Data Foundationshipped (MVP): Rocky v26r1 PDF ingest, manifest metadata, vector search + payload filters on Qdrant. Planned: Unstructured.io, BM25 fusion, structured material API.
  2. Trusted Agent Layershipped in prod: LangGraph routing, mandatory citations, three DEM helper tools, Pydantic validation, eval suite (14 golden cases), Streamlit UI with SSE streaming. The engineer remains orchestrator.
  3. Solver Orchestrationdeferred, design complete: async CAE jobs via Docker + queue; human approves parsed results. No autonomous FEA.

Commodity AI: the orchestration and data boundaries are the product — LLM provider is model-agnostic (OpenRouter, on-prem Qwen/YandexGPT-class via API).

Live demo: cae-copilot.dmitryivanov.dev


What (shipped — Stage 2)

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

Demo paths (one request takes one branch):

  • Knowledge / setup / scripting: Question → RAG → Citation → Report
  • Calculation: Question → DEM tool → Report

See architecture overview · Data-First.


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

What it is not

  • Solver automation or commercial CAE integration (Stage 3 deferred)
  • Autonomous engineering decision making
  • AS-IS click automation or "2× faster in ANSYS"
  • Smart chatbot / "chat with your PDF" only

Full scope: problem-statement.md · requirements.


Resume bullet (EN)

Designed a 3-stage enterprise AI adoption roadmap for CAE knowledge workflows: Rocky PDF data foundation (vector RAG + metadata filters), LangGraph agent layer with Pydantic-validated DEM tools and SSE streaming UI, and async solver orchestration (deferred). Addresses "AI on chaos" and black-box trust via Data-First split and human-orchestrator HITL.