Portfolio
Selected work
Two projects: an AI certification product in open beta, and an MIT-licensed task board made of folders.
01 · R&D
Certestic
Practice exams written for what you are studying, then a report that says where the points went.
- A product taken end to end: positioning, catalogue, credit model, exam runtime, deployment.
- AI only where it changes the product, with the cost and the failures designed in.
- Trust here comes from the report, not the score.
- Role
- Product, design, engineering
- Model
- Open beta · credits
- Stack
- Next.js · TypeScript · GCP
The problem
A dump teaches the test, not the subject. One scanned PDF gets passed around for years until the mock score looks right. Vendor practice exams are the honest alternative, but there are few, they never change, and they cost as much as the real exam.
Nobody sells the diagnosis. A score is not information. “Kubernetes networking, three of eight correct, weakest domain three exams running” is.
Certestic goes the other way: new scenarios and new wording every run, so memorising does not work.
The product
One loop, four steps.
Scope
Pick a track, set the question count, narrow the topics in plain English.
Generate
A fresh paper written against the track's current objectives.
Sit
Countdown, navigator, review flags, no feedback until you submit.
Diagnose
Every answer mapped to a topic: what held up, what cost points, what to try next.
After two exams, reports roll up into mastery by domain and a trend across attempts.
Credits, not a subscription. Generating an exam is the only part that costs real money, so an account starts with 300 credits and no card. One plan in beta, priced as feedback: early accounts keep their credits and hear about pricing first.
The catalogue is how people arrive. Pages named the way people search — “CKA practice exam”, “AWS AI Practitioner questions” — with category pages for people who know the area but not yet the certification.
What it takes
Generation is the easy part. Each exam is written against the certification's own objectives and question range: a 15–60 question track does not produce a 200-question paper. The learner's scope has to bind what gets written, not suggest it.
Nothing reaches a screen unchecked. Every question is checked against a schema; one that does not parse, or maps to no topic, fails the check instead of the learner. The report is computed from the sitting, so the score can be defended.
Two codebases. A Next.js app for what a learner touches, and a separate API for generation, scoring and reports, so the slow part can fail without taking the exam screen with it. A credit ledger agrees with what was generated; each exam record stores the paper the learner sat.
02 · Open source
flow-trail
A task board made of folders: three directories in git, one markdown card per piece of work.
- A tool extracted from five months on a real product, not designed in the abstract.
- Rules backed by measurements, including two blind runs that changed the design.
- Distribution as design: one command, nothing to host, MIT.
- License
- MIT
- Install
- npx skills add
- Runtime
- None — files in git
The problem
A session forgets. A chat settles a design question, the tab closes, and the reasoning is gone. The next session rebuilds it from the code, which records only the option that won — and sometimes gets it wrong.
A ticket tracker records what is left to do, not why the plan looks like this. A wiki goes stale as soon as the work moves. I wanted a record that lives beside the code and outlives the tool that wrote it.
The idea
Three folders — backlog/, doing/, done/ — and one markdown card per piece of work. The folder is the status, so nothing falls out of date. A move is git mv, so the card keeps its history and git log --follow still finds it.
Nothing runs. No server, no database, no token. That keeps the record readable in ten years, and lets an agent and a person work the same files.
- README.md
an idea: the problem, the trigger, why it waits
git mv → doing/
- README.md
- findings.md
- progress.md
the plan: phases, a baseline, a next action
git mv → done/
- README.md
- findings.md
the record: what changed, the numbers, what lost
closed; history kept
What a card holds
Six headings, fixed order: Outcome, Contract, Decisions, Lessons, Residual, Where the proof lives. Closing is mechanical: conclusions move up, evidence stays, the file that held state is deleted.
A decision keeps the option that lost. Code records only the winner, so the row is written that session. Each phase carries a Verify command with its output underneath — [v] means written, not verified, which is the usual state.
What the measurements changed
A rule that could not be a grep or a test was deleted, not turned into a tool. The README states a measurement beside each rule, and a script checks the docs.
| Measurement | Result |
|---|---|
| Mandatory closing phases completed, across 7 plans | 0 of 7 |
| Closing phases cut after the measurement | 4 |
| Archive size, 10 closed cards, before → after | 9,003 → 1,247 lines |
| Blind runs of the workflow, both graded | 2, both changed the design |
Trying it
Other tools do adjacent jobs: Superpowers holds the process, OpenSpec the contract, Beads the queue. This holds the reason. One command installs it, and the first card writes the AGENTS.md pointer so the next session finds it.
$ npx skills add benxgao/flow-trail -g- plan the login refactor
- a card in doing/: baseline measured, phases, the first one named
- where were we?
- the next line, the phase, and the decision behind it
- close this card
- README and findings distilled, progress.md gone, git mv to done/
Try one card. If you never come back to it, it cost three files.
Reach for it when
- a choice has to be explained months later
- work pauses in one chat and resumes in another
- a number has to hold up when its command is re-run
Skip it when
- the change is one commit
- one session finishes it
- you want a hosted board with a UI
Contact
A system that needs building, fixing or untangling? Send me a message.