Tuan Nguyen

Personal AI portfolio / September 2026

Learning by doing.
Building with AI.

I’m Tuan, an accountant turning everyday problems into practical AI systems. This is a collection of what I’ve built, tried, and learned along the way.

Explore the work

01 / Selected work

From useful ideas
to working routines.

Personal systems, professional applications,
and experiments worth learning from.

02Applied at work

Accounting / Human review

AI inside the
audit workflow.

I use enterprise AI tools to organize journal entries, compare process narratives, and surface inconsistencies for review.

Enterprise AICopilotSharePoint
Read the project

What I redesigned

Repeatable tasks across journal-entry classification, walkthrough preparation, and control-documentation review. AI applies prior criteria, drafts updates, and produces exception lists for the team to assess.

A specific design choice

Folder-scoped SharePoint agents narrow the evidence to the task. A broader parent-folder agent helps discover missing support when the focused context is insufficient.

The review standard

I check suggestions against underlying support before using them. AI helps prepare and organize work; professional judgment stays with the audit team.

What I learned

Context selection and exception handling matter as much as the prompt. This summary describes general methods; it contains no client data or workpapers.

03Recurring personal workflow

Knowledge / Workflow design

A knowledge system
that keeps my voice.

I built AI-assisted reading, journaling, and habit workflows with clear instructions for organizing records while preserving my own writing.

ClaudeObsidianPersistent instructions
Read the project

What I built

Reading logs, Kindle-highlight parsing, vocabulary records, linked ideas, monthly journals, and habit trackers. Raw notes become structured records; later reviews look for patterns across them.

Rules that make it usable

Preserve my spelling and phrasing. Keep AI commentary visibly separate. Confirm activity dates. Record what happened rather than treating plans as completed activities.

What changed

I can spend more effort reading and reflecting, with less effort formatting and maintaining records.

What I learned

Persistent instructions are part of the product. Preserving authorship and distinguishing source writing from AI interpretation are deliberate design choices.

04Built & tested

Data reliability / Agent boundaries

Workout logging,
with a clear handoff.

I migrated a health project to Obsidian and GitHub, separated workout transcription from planning, and added checks for consistent records.

GitHubPython validationApple Health exports
Read the project

The friction

Typing between sets interrupts a workout. Inconsistent connector data also makes later analysis harder to trust.

What I designed

A narrow logging role for Alfred and a separate planning and review role for Claude. Python scripts rebuild derived summaries and validate the record structure.

The source check

When connected data conflicted with recorded values, I worked around it using direct Apple Health exports for completed-period verification.

What I learned

Separate transcription from judgment. Generated summaries should be reproducible. Written agent boundaries help guide behavior, but do not themselves enforce access permissions.

05Ongoing experiment

Agents / Rules & evaluation

A trading experiment
with a rulebook.

I designed and iterated a small AI-assisted trading experiment around versioned rules, scheduled roles, persistent plans, and explicit failure conditions.

ClaudeCloud tasksEvaluation design
Read the project

What I explored

How an agent workflow can derive a plan, monitor state, and operate within predefined boundaries without repeatedly improvising its rules.

What I designed

Separate scheduled roles, an approval gate for new plans, a persistent approved plan, expiry and kill conditions, a halt switch, and a recurring review of rule drift and exceptions.

The important limitation

The experiment has a small sample and has not established an advantage over buy-and-hold. Positive trades alone do not prove that the system works better.

What I learned

Automation reliability and investment performance are separate questions. A rulebook needs clear precedence, measurable constraints, and logs that make failures visible.

06Explored & refined

Infrastructure / Cost & scope

Learning what it takes
to run an agent.

I configured OpenClaw, learned terminal troubleshooting and GitHub backups, then moved to Cowork after weighing maintenance effort and usage costs.

OpenClawClaude CoworkGitHub
Read the project

What I tried

A hosted agent setup with API-based models, configuration recovery, and backups. I learned to choose model capability around task complexity and context requirements.

Why I changed course

Execution costs, hosting costs, and time spent repairing the setup made a simpler subscription-based workflow more useful for my needs.

What I carried forward

Separate project contexts, limited file access, recoverable configuration, and a habit of questioning whether each proposed fix is actually needed.

What I learned

The most autonomous setup is not automatically the most useful one. Maintenance and recovery are part of the cost of operating an AI system.

07Retired exploration

Product / Learning to code

Aestheon:
from idea to specification.

I used Claude to explore a workout app concept, define gamification, and generate Swift logic through an iterative design → spec → code loop.

ClaudeSwift / XcodeProduct design
Read the project

The idea

Help people identify gym equipment and exercises from photos when they do not know the names to log.

What I produced

A product roadmap, a Greek-mythology-themed gamification specification, a linked knowledge base, and Swift logic for experience points and level progression.

Where it ended

I retired Aestheon on September 14, 2026 because the original pain point had been solved another way. These were exploratory artifacts, not a launched app.

What I learned

AI let me work across unfamiliar product and coding tasks. Reassessing whether the problem still needs solving is part of building, too.

08 / Creative side quest

Meet Tally.

A custom animated saola companion for Codex, created with AI-assisted artwork and packaged as a v2 sprite atlas. A small project in character design, animation states, and visual quality checks.

AI artworkAnimated spritesCodex pet

Click Tally.

Curiosity, with a field ledger.

02 / How I build

The controls are
part of the work.

01

Start with the source.

Use clear records and explicit instructions. Verify what an integration returns when it conflicts with the underlying data.

02

Make the roles clear.

Decide what the agent can do, what needs judgment, and how it should handle uncertainty or missing information.

03

Keep learning visible.

Document failures and tradeoffs alongside wins. Refine a useful routine, and retire an idea when the problem moves on.

03 / About me

An accountant.
A curious builder.

I’m a California CPA with a background in public-company assurance. I bring the same questions to AI that I bring to accounting: Where did this information come from? Does it reconcile? What happens when something goes wrong?

My projects span professional workflows and personal experiments. I’m learning by building systems I can actually use—and by being honest about what they can and cannot do.