All work

Qualifi.ai

Independent project · Pre-launchAI product

A career coach that remembers, and says what it doesn’t know.

An independent project: an AI career coach, Compass, that keeps your goal, plan and résumé in view across a whole job search, treats “I’m not sure yet” as a real answer, and shows its reasons. Designed and directed by me, built with AI coding agents, in pre-launch.

Qualifi’s North Star weekly plan on a monitor against a grey studio backdrop.
Role
Designed and built by me
Timeline
First Figma concept in 2024. Rebuilt as a working product, May to August 2026.
Year
2024 — 2026
Status

Pre-launch · demo build

The 2024 concept on Behance (opens in a new tab)
Tools
Figma · Next.js · Vercel AI Gateway · Claude +3
Type
AI product · Web & native

00 · TL;DR

The thirty-second version.

Six lines on Qualifi.ai. The chapters below show the work behind each one.

Role
Independent project, designed and built by me. I set the direction, designed the product and its AI behaviour, and directed the build. The code was written by AI coding agents working to my briefs and reviews.
Team
Me, plus AI agents in strategy, build, audit and review lanes.
Timeline
First Figma concept in 2024. Rebuilt as a working product, May to August 2026.
Scope
An AI career coach for people who don’t know what’s next: web app, iOS and Android shells, 13 product surfaces, an AI eval suite, and a key-free demo mode.
What I decided
Direction before applications. One coach with visible, editable memory instead of a stack of tools. “I’m not sure yet” as a first-class answer. A non-AI fallback for every AI feature. Evals written as design rules, then used as runtime guardrails.
Outcome
Pre-launch, so no usage numbers yet. What exists: 13 working surfaces, a key-free demo, and voice rules enforced by evals and runtime grounding checks — no generated output reaches a user with a number or quote they didn’t give.

Chapter 01 of 09 · The problem

A job search is ten tools that don’t talk to each other.

When people lose a job, or quietly decide they’re done with one, they reach for tools: a résumé builder, a job board, a tracker spreadsheet, a chatbot for cover letters. Each is fine alone. None knows what the others know, so the person becomes the integration layer and re-explains themselves at every step.

They also assume you already know what you want, and help you apply faster. For someone stalled or just laid off, speed is the wrong help.

01Trust is the product
One invented fact — a fake job, a made-up metric in a résumé bullet, a claim to remember something it doesn’t — spends trust that doesn’t come back.
02AI cost is real
A free tier on an expensive model with a “generate again” button loses money on every user.
03AI fails
Providers go down, keys go missing, outputs break schema. The product can’t go blank when that happens.

Chapter 02 of 09 · Where it started

The 2024 concept looked complete and solved the wrong problem.

I first designed Qualifi in 2024 as a Figma concept: a dashboard that pulled skills, courses and mentor matching onto one screen. It was still a stack of features, just on one page.

When I came back to it in 2026 the question changed, from “how do I put the tools in one place?” to “what would it take for one coach to actually know you?”

FIG. 002Where it started: a dashboard of features. Ratings, courses and enrolment numbers in the artwork were illustrative.

Crossroads on desktop and Compass on a phone, from the key-free demo build. The person and companies are fictional fixtures.

Chapter 03 of 09 · The people

Three situations, three registers.

I designed for three situations, which became three tones Compass has to handle. The second group drove the most design work: they are the worst served by job tools, because they don’t have a target to optimise for yet.

  • 01

    Just laid off

    High pressure, low patience.

    NeedsOne concrete move this week, not a framework.

  • 02Drove the most design work

    Employed but done

    Not urgent, quietly stuck.

    NeedsPermission to explore, and a cheap way to test options.

  • 03

    Early, or checking in

    Lower stakes.

    NeedsA sense of direction more than tactics.

Chapter 04 of 09 · The wedge

One coach that remembers.

The core decision was to make the coach the product, and treat everything else as places the coach can work. Compass is the coach you talk to. Crossroads is where you compare directions when you don’t have one. North Star is the 13-week goal you commit to. Recap shows how far you’ve moved.

The résumé builder, outreach drafter, application co-pilot and interview Prep Room all read the same record, so nothing starts from zero.

Surfaces

  • CrossroadsCompare directions
  • North StarCommit to 13 weeks
  • RecapSee what moved

CompassThe coach you talk to

  • Goal
  • Plan
  • Résumé highlights

Tools that read the same record

  • Résumé
  • Outreach
  • Applications
  • Prep Room
  • Jobs
FIG. 003The wedge: the coach is the product, and every other surface is a place the coach can work.

What I cut to protect it

  • Auto-applyThe most requested feature in the category, and the opposite of the point. “We never auto-apply” is on the pricing page as a promise.
  • Mentor and recruiter sidesDesigned, partly built, then flagged off. A two-sided marketplace is a second company.
  • Community feedHidden for launch.

Chapter 05 of 09 · AI UX patterns

Nine patterns for an AI people can trust with a career.

Most of the design decisions were about how the AI behaves, not how the screens look. I wrote each behaviour down as a rule, designed the states that show it, then tested it. Four are open below; the other five open from the index.

Decision 01P1 · Memory and consent

Memory you can read, check and edit.

“Compass remembers you” is the pitch. The design question was what “remembers” should mean so that it is both useful and true. I chose structured memory over chat recall: Compass keeps your goal, plan and résumé highlights and re-reads them at the start of every conversation. It does not store past chats as memories.

So I made it visible: one page shows exactly what the model is given each turn, and says its own limit in plain words. Memory then shows up in the product: the first message in a thread is built from what you already said.

The trade-off

I gave up the “it remembers everything” headline. Better to under-claim and have the product feel more knowledgeable than promised than the reverse.

Screens and annotations
  1. Re-read at the start of every conversationThe count line says how much is held and when it is read.
  2. Exactly what the model is given each turnNo hidden profile. If it isn’t on this page, Compass doesn’t know it.
  3. The limit, said plainly“Not a record of every chat.” The page admits what memory isn’t.
  4. Edit or remove it at the sourceEach item links to the place you would change it.
FIG. P1.1What Compass remembers: the whole of the model’s memory, on one page.
  1. Opens with what you already said“I read what you told Compass…” — no second intake form.
  2. An offer, not a pushPaths to compare are suggested, and only run if you say so.
  3. Next steps from the planSuggestions come from the record, not a generic prompt list.
FIG. P1.2Memory in use: the first message of a thread, built from onboarding answers.
  1. Your data stays yoursNever used to train a model, never shown to recruiters unless you choose.
  2. Take it all with youOne JSON file of everything Qualifi holds.
FIG. P1.3Consent sits next to memory: export everything, delete everything, never used to train a model.

Decision 02P2 · Uncertainty

“I’m not sure yet” is the first answer, not the last.

The second question in onboarding is “What do you want from the next move?” The first option is “I’m not sure yet, help me figure it out.” Putting it first was deliberate: most products bury “not sure” at the bottom, or leave it out, because it doesn’t map to a funnel.

Choosing it branches onboarding into a short clarity exercise — a time you felt good at work, in your own words; what gives you energy; three one-line sketches of possible futures — and it ends at Crossroads instead of a job list. Elsewhere, uncertainty is labelled: the interview rubric shows dashes before your first answer, not zeros, and empty states say what’s missing instead of faking progress.

Screens and annotations
  1. First option, not last“I’m not sure yet — help me figure it out.”
  2. No wrong answer, including noneThe helper line gives permission before the options do.
  3. Compass says what it’s holdingThe header counts the answers it has, so the quiz feels like it is going somewhere.
FIG. P2.1Onboarding, step 3: the honest option comes first.

The not-sure branch, in five states

  1. 01 Not sure

    Chosen first, branches the flow.

  2. 02 A good day

    In your own words, not a picker.

  3. 03 Three futures

    One line each. Rough is fine.

  4. 04 That’s a start

    Ends at Crossroads, not a job list.

  5. 05 Three paths

    The first of three, each cheap to test.

FIG. P2.2The branch from “not sure” to three testable paths. Each answer gets a one-line reply from Compass.

Decision 03P3 · Explainability

If the AI makes a judgement, it shows its reasons.

Crossroads turns the clarity answers into two or three directions. Each path has a name, why it fits you — quoting your own sentence back — what the day-to-day is really like, including the unglamorous part, an honest trade-off, and the smallest real experiment to test it in one to three weeks. A conversation to have, a thing to build, a day to shadow. Not “research the field.”

Jobs follow the same rule. A fit score is labelled provisional until there’s a résumé to compare against, and says why. Every listing shows where it came from, and a ghost-job score explains itself on hover. Signal Mode hides likely ghosts by default and tells you how many it hid.

Screens and annotations
  1. Quotes your own wordsWhy it fits is argued from what you said, verbatim.
  2. The unglamorous partDay-to-day reality, including what people don’t enjoy.
  3. An honest trade-offWhat you’d give up, named before you commit.
  4. Smallest real test, with a timeframe“Test it in ~10 days”: a real action, not more research.
FIG. P3.1One Crossroads path, top to bottom. The reasons are the product.
  1. The ghost score explains itselfPosted 78 days ago; generic “always looking” wording.
  2. Source shownEvery listing says where it came from.
  3. Provisional, until there’s a résuméThe fit is a read, labelled as one.
  4. Why it’s provisional“Built from your plan — add a resume to sharpen this read.”
FIG. P3.2A job the coach is unsure about says so — and says why.

Decision 04P4 · Failure modes

Designed for the AI failing, because it will.

Every AI feature has a deterministic fallback. With no key, an 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, never a blank screen.

Refusals are designed too. Anything from the user’s data — a pasted job post, a résumé — is wrapped and treated as data. Asked to reveal its instructions, Compass declines in one line and goes back to the person’s goal. No lecture.

Screens and annotations
  1. Treated as data, not an instruction
  2. Declines in one lineNo explanation of the rules, nothing to argue with.
  3. Back to the person’s goal“…what were you actually trying to work out?”
FIG. P4.1A prompt-injection attempt, and the whole of the reply.
  1. Request
  2. Model
  3. Grounding checkNumbers and quotes traced to the person’s own input
  4. PassShown to the person
Fail, outage or no key

Deterministic fallbackRule-based, built from their real inputs

→ shown to the person
  • No blank screens
  • No invented numbers or quotes
  • Same check as the eval
FIG. P4.2Every generation passes a grounding check. A failure is not an error state; it is the fallback.
Prompt injection
Pasted job posts and résumés are wrapped as data. Requests for its instructions get one line, then back to the goal.
Integrity
It will not invent accomplishments. No story for a question? It helps you find a real adjacent one.
Real-world limits
Legal, medical, immigration and financial questions go to a licensed professional.
Crisis language
Care and a crisis line first, before anything career-related.
FIG. P4.3The other edges, each with a designed response.

Decision 05P5 · Human control of drafts

Compass drafts. People send.

The outreach engine writes an email and a LinkedIn message, then gives you Warmer, Tighter and More specific to steer it. Nothing leaves Qualifi until you copy it yourself, and status — sent, replied, no response — is something you mark, not something it assumes.

Before Compass writes any artifact it asks for the one or two specifics that make it real, like the posting or an actual win. The résumé builder asks for real details first (“Even a rough sentence helps”), and everything it writes is editable section by section. A generic cover letter teaches people that volume is fine.

Screens and annotations
  1. The next step, in plain words“Read it once in your own voice… copy the email, send it, and mark it sent.”
  2. Steer, don’t regenerate blindlyWarmer, Tighter, More specific. The draft updates in place; nothing is lost.
  3. You copy, you sendQualifi never sends on your behalf.
  4. You mark the statusSent, replied, no response: recorded by you, never assumed.
FIG. P5.1An outreach draft. Every control that matters belongs to the person.

Decision 06P6 · Tone

A written voice, treated like a design spec.

Compass is short by default, takes a position, and disagrees when it’s right to — if a plan is weak, it says so in the first two sentences, before any encouragement. No filler openings, no hype words, no emoji. Outreach has its own peer-tone spec: 60 to 120 words, one specific reason, one ask, and a banned list of sales-email phrases.

Tone also adapts to the person, and the person controls it. The “why now” question in onboarding sets the register: someone just laid off gets a steady, action-first first reply. A Career season setting lets them change that later, and it is never shown to recruiters.

Screens and annotations

compass / voice.mdSpec

Short by default
Two to four sentences unless the question earns more.
Take a position
“It depends” is not an answer. Name the branches and say which way you’d lean.
Disagree when it’s right to
If a plan is weak, say so in the first two sentences, before any encouragement.
Ask before drafting
Get the one or two specifics that make an artifact real.

Never says

  • Great question
  • I hope this helps
  • journey
  • unlock your potential
  • I hope this email finds you well
  • circling back
  • quick question
FIG. P6.1The voice rules, as written for the model and for reviewers.

From “why now” to the first reply

  1. 01 Why now

    Sets the register.

  2. 02 First reply

    Steady and action-first after a layoff.

  3. 03 Career season

    Change or clear it any time.

  1. Why-now sets the register
  2. “Do not start applying for a week.”
  3. Never shown to recruiters
FIG. P6.2The register is set by the person, used by the coach, and changeable later.

Decision 07P7 · Streaming and latency

Waiting is a state worth designing.

Compass replies stream in as they’re written, with a stop button available the whole time. Before the first token, a thinking line says what it is doing in its own voice — “Thinking about which questions are worth asking…”, “Pulling up where you said you wanted to land…” — picked from small pools grouped by what’s happening, and seeded so it doesn’t flicker on re-render.

Tool calls show as a status line. Long generations, like Crossroads and the plan, get a named loading state that says what is being read.

Screens and annotations

One reply, three states

  1. 01 Thinking

    Says what it’s doing, in its own voice.

  2. 02 Streaming

    Words arrive as written. Stop stays in reach.

  3. 03 Done

    Suggestions return; the humility line stays.

  1. Says what it’s doing
  2. Streams as it writes
  3. Stop is always there
  4. The humility line
FIG. P7.1The same thread a second apart: thinking, streaming, done. The streamed text is the eval-tested no-key fallback reply.

Decision 08P8 · Evals as a design tool

Voice rules became tests, then runtime guardrails.

The voice rules weren’t just a style guide. The eval suite checks Compass against them: it takes a position on open questions, doesn’t fabricate, handles a layoff in the right register, stays brief on simple questions, resists injection (including role-play variants) and never uses the banned words. Fifteen deterministic cases run on every change with no API key; fifteen more run against the live model when a key is present.

The most useful step was promoting some of those checks into runtime guardrails. Interview feedback, résumé content, recaps and plans pass the same grounding checks before anyone sees them: if the model introduces a number the person never gave, or “quotes” something they never said, the output is rejected and the grounded fallback is used. The eval and the product enforce the same definition of honest.

The trade-off

Some good model outputs get thrown away by a strict check. I accepted that: a slightly plainer answer is better than a confident fabrication.

Screens and annotations
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
FIG. P8.1Rule → eval → runtime. 15 deterministic cases gate every change; 15 more run against the live model before release.

Scoring that teaches

  1. 01 Before

    Dashes, not zeros, until there’s an answer.

  2. 02 After

    What worked first, then what to tighten.

  1. Dashes before the first answer
  2. Each dimension asked as a question
  3. What worked first, then what to tighten
  4. Updates after every answer
FIG. P8.2The Prep Room rubric before and after an answer. Each dimension is asked as a question, so the score teaches.
FIG. P8.3The Prep Room, 24 seconds, no sound. Demo mode with deterministic scoring, not a live model call.
Read it instead — what happens in the recording

An answer about rebuilding an ignored churn dashboard is typed and submitted. Compass reads it (“Compass is reading your answer…”) while the rubric holds its old values, then replies with what worked, what to tighten and a stronger version. The rubric on the right moves: clarity, specificity, confidence and structure each update out of 10. The session ends in a debrief that names what landed and what to sharpen next.

Decision 09P9 · Cost and model routing

One door for every model, and a budget on each feature.

All model calls go through one door, the Vercel AI Gateway: one budget cap, provider failover (Anthropic first, OpenAI as fallback) and one place to see spend. Paying users get the strongest model for coaching, plans and résumés; free users get a strong but cheaper one. Free traffic is most of the volume, so this is the biggest margin lever.

Prompt caching and model routing keep a free tier affordable. A test fails the build if an expensive model becomes the default by accident.

Screens and annotations

Surfaces

  • Compass
  • Crossroads
  • Plans
  • Résumé
  • Prep Room
  • Outreach

One door

AI Gateway
  • Budget cap
  • Provider failover
  • Cached system prompt

Anthropic first · OpenAI on failover

Model lanes

  • PremiumPro: coaching, plans, résumés
  • DefaultFree tier: most of the traffic
  • FastSmall jobs: labels, short checks
  • Tier-aware routing
  • Cached system prompt
  • Per-feature budgets
FIG. P9.1Model routing. Free users get a monthly allowance of coach messages and one Crossroads exploration; regenerating paths is a Pro feature.

Chapter 06 of 09 · Mobile

Same product, phone first.

The mobile build is not a squeezed desktop. A five-item tab bar replaces the sidebar, the composer and the humility line stay above it, and long pages such as Crossroads and North Star are written to be read with a thumb.

  1. A native tab bar, five destinations
  2. Composer and humility line stay above it
  3. The goal, quoted, always at the top
FIG. 004Compass, Crossroads, North Star and the Prep Room at 390 px.

Chapter 07 of 09 · The system

Four surfaces, one weekly rhythm.

Compass is the coach you talk to. North Star is the plan you committed to. Crossroads is the fork where you decide which plan to make. Recap is the mirror. North Star breaks a 13-week goal into three outcomes a week, and a Sunday review looks at what landed without guilt: “A quiet week — that’s data too, not a verdict.”

  • North Star

    A quoted 13-week goal, three outcomes this week.

  • Sunday review

    What landed, without guilt.

  • Recap

    Your own words, quoted back exactly.

FIG. 005The weekly rhythm: commit, review, look back.

Chapter 08 of 09 · Outcomes

What’s true today.

Built
A complete product across web and native: 13 surfaces, a working demo mode, an eval suite and runtime grounding checks.
Not yet
Public launch, paying users, retention data.
Users
None to report. No metric on this page is a usage number.
First metric I’d watch
Activated sign-ups by source in the first two weeks, where “activated” means finishing onboarding and starting a Crossroads or North Star.

Chapter 09 of 09 · Reflection

Start from the conversation, design screens second.

What I’d do earlier
I started in 2024 with 41 Figma screens of a dashboard. The product only got good when I started from the conversation and asked what the coach needed to know.
What I got right
Writing the AI’s behaviour as rules, then tests, then runtime checks. It turned “make it sound good” into something I could review and enforce.
A trade-off I’d revisit
The demo uses a scripted router so it works without keys. That trade-off shows: it answers some prompts in ways the real coach never would. Next, I’m routing the demo through the same eval-tested fallbacks.

What I’d do next

  1. 01Launch to a small group and learn whether people come back for the Sunday review. That’s the habit the whole product depends on.
  2. 02Episodic memory, with consent: let people choose specific conversations for Compass to remember, and show them on the memory page.
  3. 03Run the model evals on every release, and publish the voice rules.
Under the hoodPlatforms, tokens, and how I worked with agents
  • Web, iOS and AndroidOne codebase: a Next.js web app, plus a native shell with Face ID lock, push and the native share sheet. Built, not yet in the stores.
  • Design tokensColour, type, radius, elevation and motion live in one theme file, with one five-tone status vocabulary across the product. Motion is transform and opacity only, and reduced-motion safe.
  • How I worked with agentsLike a small team: I wrote the strategy and the briefs, agents built in parallel lanes, separate agents reviewed, and nothing merged without the build, the evals and my sign-off.

What shipped

  • One coach with visible, editable memory across 13 surfaces.
  • “I’m not sure yet” as a first-class path, ending in real one-to-three-week experiments.
  • Every AI feature has a grounded, non-AI fallback; evals double as runtime guardrails.
  • Web, iOS and Android from one codebase, with a key-free demo mode.