Traceability over fluency
A persuasive answer is not evidence. Inputs, transformations, tests, and sources must remain inspectable.
Consulting scope · Curriculum draft · Human-supervised AI
Build engineering agents that expose their work.
I help Energy R&D teams structure bounded agents around approved MATLAB and Simulink tools, deterministic tests, visible evidence, explicit stop conditions, and engineer approval.
Execution contract
Requirement → Plan → Tool → Test → Evidence → Approval
The agent may accelerate analysis and documentation. It does not own the engineering decision.
Syllabus
Seven modules from architecture to accountable output.
01Design bounded engineering agents
Define the objective, permitted actions, unavailable information, stop conditions, and approval authority.
02Structure requirements and technical context
Convert source documents and engineering questions into explicit, traceable working context.
03Invoke MATLAB functions safely
Use typed inputs, unit declarations, output schemas, deterministic scripts, and recorded tool results.
04Inspect Simulink architecture
Read model hierarchy, interfaces, parameters, signal paths, and diagnostics without inventing model intent.
05Verify calculations and artifacts
Run tests, compare tolerances, identify missing evidence, and preserve failed results for review.
06Generate engineering evidence packages
Assemble assumptions, tool calls, outputs, plots, tests, limitations, and unresolved questions.
07Escalate and approve
Stop on missing inputs, failed tests, unsupported assumptions, safety relevance, or unclear authority.
Worked transcript
A bounded, inspectable pilot for the workflow.
Example artifact
Battery pack safety review flow
Proposed by engine: revise current window, update thermal guard-band, and recalculate hotspot trend for a conservative start-up schedule.
Tool output: checks identified two edge cases outside acceptance bounds and retained raw data, scripts, and assumptions in the evidence set.
Engineer action: denied automatic acceptance, added revised sensor placement and boundary conditions, and routed the change for a second validation run.
Outcome: no safety decision was executed without human approval; every step stayed replayable and auditable.
Operating principles
Tools over imitation
Use approved engineering tools for calculations and model inspection instead of asking a language model to approximate them.
Failure remains visible
Do not suppress failed checks, missing files, unsupported assumptions, or uncertainty to produce a cleaner narrative.
Approval stays human
Safety, compliance, release, and acceptance decisions remain with authorized engineers.
Inspectable references
Use tested artifacts as the pattern, not generated claims.
The MATLAB Simulink Energy Lab exposes model checks and release evidence; its focused battery thermal case study keeps equations and assumptions inspectable; VoltRL binds protocol, results, and data provenance; and the BESS QA/QC Toolkit turns status, completeness, deadlines, traceability, and acceptance into reviewable gates. The full open engineering portfolio keeps failed assumptions and engineering boundaries visible.
Bring one bounded workflow, approved toolchain, and review bottleneck.Discuss an agentic engineering workflowInspect the validated model repository