AI WORKFLOW ENGINEERING -- AGENTIC SYSTEMS
Embedded systems engineer (C++/Qt) turned AI-workflow builder: multi-agent pipelines, MCP integrations, eval gates, and the enablement work that takes a team from zero to daily AI use.
AI WORKFLOW ENGINEERING -- AGENTIC SYSTEMS
Embedded systems engineer (C++/Qt) turned AI-workflow builder: multi-agent pipelines, MCP integrations, eval gates, and the enablement work that takes a team from zero to daily AI use.
01 — Workspace
Position-addressed, harness-agnostic -- the knowledge architecture I build for teams, running my own life. Browse it. README.md is open for you.
02 — Capabilities
Multi-step LLM pipelines with tool use, triage cascades, and human-in-the-loop gates -- built to run unattended and fail loudly.
Position-addressed knowledge bases, retrieval indexes, and CLAUDE.md/CONTEXT.md routing that any LLM harness can navigate cold.
Hard verification gates, automated test generation, requirement-to-test mapping, and refactoring pipelines that cannot ship silent breakage.
Took a production C++ team from 0% to daily AI use: workshops, live demos, custom skills, and SOPs that survive after the champion leaves the room.
Custom agents, skills, hooks, and MCP integrations; Python pipeline tooling that turns one-off AI wins into repeatable team infrastructure.
4+ years of production C++/Qt on John Deere displays, 100+ defects resolved -- the engineering rigor underneath the AI work.
03 — Pipeline
raw sources -> one inbox
skip / light / deep routing
structured claims, timestamp anchors
hard verification -- fails loudly
cross-linked synthesis layer
humans + agents query one index
raw sources -> one inbox
skip / light / deep routing
structured claims, timestamp anchors
hard verification -- fails loudly
cross-linked synthesis layer
humans + agents query one index
feedback -> recompile04 — Runbook
05 — Human
Embedded systems engineer at John Deere by day. Building, ideaverse-os, yt-vision-pro by night. I build the AI tooling I wish I had at my engineering workbench, then drive adoption across teams that didn't know they needed it.
POSITION OVER NAMES
Folder structure is the API.
Position-addressed memory beats named imports. Names rot when you rename things; positions don't. The whole site is built on this idea — your folder paths ARE your routing, your taxonomy, your contract. ICM makes this explicit.
06 — Contact
If you are hiring for AI workflow or agentic-systems work, the resume is the 30-second version and my agent can answer the rest.
or email kevtrinhnguyen@gmail.com
I take engineering teams from "AI is a demo" to "AI is how we work": multi-agent pipelines with eval gates, knowledge systems both humans and agents can navigate, and the enablement that makes it stick.
Evidence: a 7-layer agentic refactor pipeline on shipped C++ at John Deere, a team knowledge hub covering 30+ vehicle models, 0% -> daily Copilot adoption, and ideaverse-os (open source).
Foundation: 4+ years production embedded C++/Qt. Austin, TX / remote.
Position-addressed: folder = type, prefix = priority. New work earns a folder slot; old work doesn't get to take up visual real estate. Rules and identity for all AI agents live in AGENTS.md (canonical); CLAUDE.md and copilot-instructions.md are 1-line shims pointing back to it.