John Yoon

John Yoon

I build reliable AI systems and write about the human judgment, Christian formation, and mission they should serve.

John Yoon

Writing

All writing →

Work

GitHub →

DJ Copilot

Citation-first AI research interface over an approved public corpus, with read-only retrieval, canonical source links, and explicit safety boundaries.

Social Gen

AI social-content generation workflow built around durable job state, asynchronous processing, and staged copy/image generation.

Hanuri Missions

Bilingual mission-support product work presented through selected stakeholder artifacts while keeping the private implementation and sensitive ministry details closed.

symlink-logger

FUSE-backed CLI for safely retiring directory symlinks by logging compatibility-path access.

aicommit-split

Deterministically split working-tree changes into semantic commit groups before running aicommit.

research-workbench

CLI tools for organizing research projects, sources, and Readwise workflows.

Technical background & project notes
  • Platform engineering, cloud architecture, and production reliability
  • Agentic AI, evaluation, guardrails, and reusable developer platforms
  • Technical leadership, mentoring, and mission-driven technology

Capability → evidence

Follow the claim to the work.

Read across each focus area to inspect projects, case studies, and writing. These are evidence types, not proficiency scores; confidential case studies illustrate patterns rather than independently verified deployments.

Reliable platforms and systems
Agentic AI, evaluation, and guardrails
Leadership and mission-oriented technology

DJ Copilot

Status: Not assessed

  1. Experimental
  2. Shipped
  3. Adopted
  4. Maintained

Release, adoption, and maintenance status have not been verified. A public link alone does not establish maturity.

Next evidence milestone (proposed): Record a dated release and the scope of public evaluation.

Inspect DJ Copilot evidence

Assessment recorded . Categories are not cumulative or percent complete.

Social Gen

Status: Not assessed

  1. Experimental
  2. Shipped
  3. Adopted
  4. Maintained

Release, adoption, and maintenance status have not been verified. A public link alone does not establish maturity.

Next evidence milestone (proposed): Publish a dated, public-safe example of failure recovery.

Inspect Social Gen evidence

Assessment recorded . Categories are not cumulative or percent complete.

Hanuri Missions

Status: Experimental

  1. Experimental (current)
  2. Shipped
  3. Adopted
  4. Maintained

The portfolio describes a product prototype; adoption is not claimed.

Next evidence milestone (proposed): Confirm which stakeholder artifacts are approved for public release.

Inspect Hanuri Missions evidence

Assessment recorded . Categories are not cumulative or percent complete.

symlink-logger

Status: Not assessed

  1. Experimental
  2. Shipped
  3. Adopted
  4. Maintained

Release, adoption, and maintenance status have not been verified. A public link alone does not establish maturity.

Next evidence milestone (proposed): Link a dated release and a reproducible usage example.

Inspect symlink-logger evidence

Assessment recorded . Categories are not cumulative or percent complete.

aicommit-split

Status: Not assessed

  1. Experimental
  2. Shipped
  3. Adopted
  4. Maintained

Release, adoption, and maintenance status have not been verified. A public link alone does not establish maturity.

Next evidence milestone (proposed): Link a dated release and a reproducible usage example.

Inspect aicommit-split evidence

Assessment recorded . Categories are not cumulative or percent complete.

research-workbench

Status: Not assessed

  1. Experimental
  2. Shipped
  3. Adopted
  4. Maintained

Release, adoption, and maintenance status have not been verified. A public link alone does not establish maturity.

Next evidence milestone (proposed): Link a dated release and a reproducible research workflow.

Inspect research-workbench evidence

Assessment recorded . Categories are not cumulative or percent complete.

activitywatch-calendar-export

Status: Not assessed

  1. Experimental
  2. Shipped
  3. Adopted
  4. Maintained

Release, adoption, and maintenance status have not been verified. A public link alone does not establish maturity.

Next evidence milestone (proposed): Link a dated release and a privacy-safe export example.

Inspect activitywatch-calendar-export evidence

Assessment recorded . Categories are not cumulative or percent complete.

Case studies

Intuit GenAI support orchestrationSupport automation and expert workflows at production scale.
Problem
Route complex customer-support work through GenAI systems without losing reliability, escalation control, or human trust.
Architecture
Orchestration layers around LLM calls, retrieval, guardrails, evaluation loops, expert matching, and production ML services.
Impact
Improved the path from customer question to useful answer while keeping quality, safety, and operational constraints visible.
Constraints
Proprietary systems; public summary focuses on problem framing, architecture patterns, and reliability tradeoffs.
Support orchestration: an illustrative pattern

A support request enters orchestration, passes quality checks, and can reach a human expert. Evaluation feeds back into orchestration.

Publication boundary Pattern only, derived from the existing public-safe summary. This is not an Intuit system diagram or a claim about deployed controls.

  1. 01 · User-facing boundarySupport request

    A question and its available context.

  2. Route the question
    02 · Automation boundaryOrchestration

    Coordinate retrieval and model calls.

  3. Review the candidate
    03 · Review boundaryQuality and guardrails

    Evaluate the candidate response before proceeding.

  4. Escalate when needed
    04 · Human boundaryHuman expert

    Make escalation and responsibility explicit.

Feedback: Quality and guardrails → Orchestration
Evaluation feedback

Design questions: latency, traceability, safe failure, and ownership of the expert handoff.

Public-summary abstraction reviewed ; not employer approval.
Capital One fraud/ML systemsFraud, risk, and production data systems.
Problem
Detect and mitigate risky activity in high-volume financial workflows where false positives, latency, and auditability matter.
Architecture
Model development, feature/data pipelines, experimentation, monitoring, and production decision-support systems.
Impact
Built ML capabilities for risk operations where model behavior had to be measurable, explainable, and production-ready.
Constraints
Financial-services work is confidential; public summary keeps customer, model, and platform details abstract.
Fraud and ML: an illustrative pattern

Features feed model experimentation and decision support. Monitoring returns observations to model development; people retain operational responsibility.

Publication boundary Pattern only, derived from the existing public-safe summary. No customer data, thresholds, proprietary models, or internal platform topology is shown.

  1. 01 · Data boundaryData and features

    Prepare inputs for model development.

  2. Supply features
    02 · ML boundaryModel and experiment

    Compare behavior before operational use.

  3. Provide model output
    03 · Operations boundaryDecision support

    Make model outputs inspectable by risk operations.

  4. Observe decisions
    04 · Human review boundaryMonitor and review

    Observe behavior and investigate exceptions.

Feedback: Monitor and review → Model and experiment
Evaluation and investigation feedback

Design questions: false positives, latency, auditability, and escalation of uncertain decisions. Specific controls are not asserted.

Public-summary abstraction reviewed ; not employer approval.
Open-source toolsResearch and local-first developer workflows.
Problem
Make everyday research, time analysis, terminal sessions, and filesystem migration safer to inspect and automate.
Architecture
Small CLIs and local automation around structured project state, calendar exports, tmux/session workflows, and FUSE-backed access logging.
Impact
Public code demonstrates practical systems taste: narrow interfaces, repeatable workflows, and tools built for real personal use.
Constraints
Designed for local workflows first, with privacy-preserving examples and public-safe documentation.
Local-first tools: a shared pattern

Local inputs pass through a narrow command-line interface to inspectable outputs. Inspection informs the next invocation, rather than hiding changes behind a service.

Publication boundary A common design pattern across separate tools, not a claim that these repositories form one integrated system.

  1. 01 · User-controlled boundaryLocal input

    Research state, activity records, or filesystem access.

  2. Explicit invocation
    02 · Local execution boundaryFocused CLI

    A narrow interface for one workflow.

  3. Write inspectable results
    03 · Human inspection boundaryInspectable output

    Review logs, project files, or calendar exports.

Feedback: Inspectable output → Local input
Inspect before the next run

Design questions: repeatability, privacy of local data, and recovery from failed operations. Guarantees vary by tool.

Public-summary abstraction reviewed ; not employer approval.

About

13+ years across production ML, GenAI, cloud platforms, and distributed systems. My work spans engineering, technical leadership, and responsible technology in Christian ministry and other mission-driven contexts.

Focus

Reliable AI in Practice

Production ML, agentic systems, evaluation, guardrails, platforms, and the leadership required to make them dependable.

Technology, Faith, and Human Agency

Theology, Christian formation, mission, and the human consequences of increasingly capable technology.

Experience & CV

Two lanes of work

Engineering alongside leadership, education, recognition, and mission. Exact career dates are not available in the published CV; undated entries are shown without a time scale or implied alignment.

Engineering progression

Roles in CV order (most recent first). Length and spacing do not represent tenure.

  1. Dates not publishedCapital One

    Senior Lead Machine Learning Engineer — enterprise LLM evaluation, agentic workflows, guardrails, and platform enablement.

    Author-reported CV source notes
  2. Dates not publishedIntuit

    Senior Machine Learning Engineer — production GenAI orchestration, support automation, ML systems, and mentoring.

    Author-reported CV source notes
  3. Dates not publishedNutanix

    Member of Technical Staff — automation, release infrastructure, distributed systems testing, and operational tooling.

    Author-reported CV source notes

Leadership, learning, and mission

Parallel areas of work; undated items are not aligned to engineering roles or ordered chronologically.

  1. Dates not publishedTechnical leadership and mentoring

    Developing engineers and enabling teams through platform quality and cross-team architecture.

    Author-reported CV source notes
  2. Dates not publishedTheological education

    M.Div. candidate at Southwestern Baptist Theological Seminary. The CV lists expected 2026; completion is not verified.

    Author-reported CV source notes
  3. Dates not publishedPatents and recognition

    The CV reports three U.S. patents and a Gloo.AI Best Concept Award. Identifiers and award dates are not published here.

    Author-reported CV source notes
  4. Dates not publishedMission-oriented product work

    Hanuri Missions: bilingual product communication with an explicit privacy boundary.

    Hanuri Missions portfolio note

Source notes reviewed . Review dates are not career-event dates.

CV source notes

Engineering

  • Senior Lead Machine Learning Engineer, Capital One — enterprise LLM evaluation, agentic workflows, guardrails, AWS/Kubernetes platform enablement, and cross-team architecture.
  • Senior Machine Learning Engineer, Intuit — production GenAI orchestration, AI guardrails, support automation, ML systems, and mentoring.
  • Member of Technical Staff at Nutanix — automation, release infrastructure, distributed systems testing, and operational tooling.
  • 3 U.S. patents.

Mission & leadership

  • M.Div. candidate at Southwestern Baptist Theological Seminary, expected 2026.
  • Interested in engineering leadership where platform quality, developing people, and Christian mission reinforce one another.
  • Built an open-source multi-agent platform recognized with the Gloo.AI Best Concept Award.

Publication record →