AI doesn’t make the calls.It makes room for them.

I design how AI behaves, and I direct AI agents the way I would lead a small team: written briefs, separate lanes, independent review. The building gets faster; the calls stay mine. Below is one product, start to finish.

Qualifi.ai: making an AI coachhonest by design.

Compass, the coach inside Qualifi.ai, gives career advice. One invented fact spends trust that doesn’t come back, so honesty couldn’t be a tone of voice. It had to be a rule, a test and a check in the product. Here is that chain, from brief to fallback.

TL;DR · the 30-second version

I design how AI behaves (what it remembers, when it admits uncertainty, what it refuses) and I run AI agents like a small team: written briefs, separate lanes, independent review, and the calls stay mine.

Role
Product design end to end: the strategy, the briefs, the behaviour rules and every sign-off
Product
Qualifi.ai, an independent project: an AI career coach (Compass) across 13 surfaces on web and native. Pre-launch.
Team
AI coding agents in parallel lanes, with separate agents auditing and reviewing
What I decided
Memory you can read and edit. “I’m not sure yet” as the first answer. Refusals in one line.
Evidence
15 of 15 offline evals passing on 28 Oct 2026. The 15 live-model cases need a key.
Outcome
Built, not launched. No users to report, and no number here is a usage number.
Fig. 01 · Rule → eval → runtime guardrailQualifi.ai · pattern P8
Design ruleEval caseRuntime guardrail
Take a positionOpen questions get a lean, not “it depends”Eval only
Never fabricateNo invented numbers, quotes or accomplishmentsGrounding check on feedback, résumés, recaps, plans
Meet a layoff steadilyRight register for “I just got laid off”Eval only
Brief on simple questionsLength stays inside the capEval only
Resist injectionIncluding role-play variantsUser data wrapped as data
No banned wordsHype words and sales phrases never appearOutreach length caps + banned list
Fallbacks are specificRecap and Crossroads fallbacks are non-empty and personalFallback served on any failed check

The same definition of honest runs in two places: in the eval suite on every change, and in the product before anyone sees an answer. Where a rule has no runtime column, it is held by the eval alone.

Read the full case

Five steps,from brief to fallback.

  1. 01Brief

    Trust is the product

    A fake job, a made-up metric in a résumé bullet or a claim to remember something it doesn’t: any one of them loses the person. So the brief started with what the coach must never do. That included defining what “remembers” means.

    AIBuilt from the briefs in parallel lanes, each with its own acceptance list.

    My callStructured memory over chat recall. Compass re-reads your goal, plan and résumé highlights, and a page shows exactly what it was given. I gave up the “it remembers everything” headline.

    Case · P1 Memory and consent

  2. 02Voice

    A voice written like a spec

    Short by default. Take a position. Disagree when it’s right to: if a plan is weak, say so in the first two sentences, before any encouragement. Ask for the specifics before drafting. A banned list of hype words and sales phrases.

    AIWrote replies against the rules; reviewers checked them against the same page.

    My call“It depends” is not an answer. And “I’m not sure yet” goes first in onboarding, not last, because not knowing is where most people start.

    Case · P2 Uncertainty · P6 Tone

  3. 03Evals

    The rules became tests

    Fifteen deterministic cases run on every change with no API key: a lean on open questions, the right register after a layoff, brevity, no fabrication, injection resistance including role-play, no banned words. Fifteen more run against the live model when a key is present.

    AIRan the suite on every batch. Nothing merged red.

    My callWhich behaviours deserve a test. A rule nobody can check is a hope.

    pnpm eval · 15/15 offline · 28 Oct 2026

  4. 04Guardrail

    Then the tests went into the product

    Interview feedback, résumé content, recaps and plans pass the same grounding check before anyone sees them. If the model introduces a number the person never gave, or “quotes” something they never said, the output is rejected.

    AIWired the check into each generation path.

    My callPromoting checks from the eval into runtime. The eval and the product enforce the same definition of honest.

    Case · P8 Evals as a design tool

  5. 05Fallback

    A failed check is the fallback, not an error

    No key, a provider outage or an output that fails validation: Crossroads, Recap, the plan and interview feedback fall back to a rule-based version built from the person’s real inputs. Never empty, never invented. Refusals are designed too: one line, then back to the person’s goal.

    AIBuilt the rule-based versions and kept a zero-key demo mode working on every surface.

    My callBreaking demo mode counted as the worst bug, because it is how anyone can try the product.

    Case · P4 Failure modes

Where agents go wrong,and what catches it.

The same habits on every build, Qualifi.ai and this site included. Each failure below happened, or nearly did, and each now has a check that runs before anything merges.

  1. 01

    It invents a number

    What it looks likeA plausible metric in a case study. “Increased conversions by 25%.” It reads fine and nobody can defend it.

    The check

    Every figure is listed with its source. Anything unconfirmed is flagged for me and stays off the page.

    Unsourced percentages: cut

  2. 02

    It drifts from the brief

    What it looks likeFive lanes, five slightly different ideas of the same button, spacing or headline size.

    The check

    Shared parts built first, and a written spec of tokens, layouts and pass/fail checks. Every round is captured on desktop and phone, light and dark.

    One spec, every lane

  3. 03

    It looks done and measures wrong

    What it looks likePerfect screenshots, slow page. On phones the headline waited for JavaScript before it could paint.

    The check

    A read-only performance and motion audit, run by an agent that didn’t build the page.

    Found by audit, fixed before merge

  4. 04

    Agents collide

    What it looks likeTwo lanes edit the same file; the merge quietly drops one of them.

    The check

    One worktree per lane and a written list of the files each one owns. Cross-lane needs go in the report, not in someone else’s file.

    One owner per file

  5. 05

    It leaks what’s private

    What it looks likePersonal contact details on a public page, or a client case going live before the client said yes.

    The check

    Hard rules in every brief, and visibility flags in the content: only approved work is listed, routed or linked.

    Unapproved work stays unlisted

  6. 06

    It says it’s finished

    What it looks like“Done” from the agent that did the work.

    The check

    Done means the build passes, the captures were looked at, and someone other than the builder reviewed it. Then me.

    Every lane, every round

Here’s what I built with AI.

A running log of the agents, workflows and checks behind each build.

  • A production website rebuilt end to end — custom Next.js, a public audit — built in parallel agent lanes and reviewed by a second model before every merge.

    Website · Agents · Reviews

  • An AI career coach with memory you can read and edit, an honest “I’m not sure yet” path, and voice rules enforced as tests and runtime checks.

    AI product · Evals · Fallbacks

  • A 72-second narrated film, composed in HTML and rendered to video — so changing the work is editing a file, not re-cutting a timeline.

    Motion · Video · Generated voice

  • A two-office law firm site with seven practice areas and thirteen guides, built with AI-assisted development and checked line by line against the Law Society’s marketing rules.

    Website · CMS · Compliance

  • A working HTML/CSS prototype of a group-trip concept, with real states — default, no shared fit and a changed estimate — instead of a static mockup.

    Prototype · Concept

The uxpro.ca proof page with the latest dated Lighthouse readings and a pass or fail beside each

01uxpro.caOct 2026

The stack,and what each is for.

Reached for every week. Short on purpose — a tool used once a quarter is a tool I’m not good at.

01ClaudeThinking partner, and Claude Code for builds with an agent squad.Daily · AI

What I use it for, every week

  • Task Automation
  • Research
  • Analysis & Decisions
  • Code & Development
  • Fast Prototyping
  • Testing & Validation
  • Agents & Reviews
  • Building real world apps