×

Individual Project · Scenario 18

An AI coach that makes design learners look again.

Design Thinking, Agile MVP and a pilot plan for an AI Portfolio Coaching Assistant at colorME, with a working product.

PresenterNguyen Viet Hung · Founder, colorME

CourseAI Product Design & Innovation · MB21CI11

Live prototypecolorme-crit.pages.dev

InstructorDr. Le Nhan Tam

Portfolios published by Vietnamese designers, 2025–26

01 The problem

Learners get one feedback cycle, where design skill needs three or four.

Learners

Can't say what's wrong

Feedback arrives after they stopped working on the piece. Weaker portfolio at course end.Assumption

Instructors

Same five notes, thirty times

20–40 pieces × 10–15 min ≈ 3–10 h of evening critique per assignment round.Estimate

colorME

Feedback is capped by instructor hours

Portfolio quality drives completion, referrals and upsell into longer tracks (up to 20.3M VND).Fact · colorme.vn

And this is what the year has to produce: 2025–26 portfolios published on Behance by Vietnamese designers. Portfolio quality is the outcome; it is built one critique cycle at a time.

02 Empathize proto-personas · validate in week 0

Two people, two fears.

Mai Anh, 24 · the career switcher

Marketing exec, Photoshop class in the evening, wants an in-house design job within a year.

Says“Is this okay?” · “I'll fix it later.”

ThinksMy colours are bright, so it must stand out.

DoesTries 4 fonts, posts once, rarely revises.

FeelsUnsure; exposed in public crits.

Fear of AI: generic feedback; being judged by a machine.

Anh Tuấn, 31 · the instructor

Senior agency designer teaching two evening classes.

Says“Squint at it: what do you read first?”

ThinksMost issues are the same five.

DoesCritiques at night; deep-dives a few pieces.

FeelsTired of repetition; protective of his voice.

Fear of AI: bland, over-confident feedback that undermines him.

02 Reflect · as-is journey

The pain is in the gaps: stuck, forgotten, exposed.

Brief

Make

Submit

Wait

Feedback

Move on

Doing

Listens in class

Works alone; tries many fonts

Posts one export

Starts next exercise

Hears 1–2 generic comments

Keeps v1 for portfolio

Thinking

“Seems easy.”

“Something's off, but what?”

“Hope it's okay.”

“Did he see it?”

“What exactly should I change?”

“No time to redo it.”

Feeling

Motivated

Stuck

Anxious

Forgotten

Exposed

Resigned

Speed

Feedback must land while the file is still open.

Self-critique

The skill we teach is seeing your own work. Train the eye, don't fix the piece.

Anchored

Where, what, why, tied to a rubric criterion.

Boundaries

Trust needs visible limits: no generated designs, instructor override.

**Crit** — a few pieces in depth, in public

**Move on** — before anything is revised

02 Define · Hills

HILL 1 · LEARNERA beginner design learner NEEDS TO know, the same evening, what to change and why, BECAUSE feedback days later arrives after they stopped working on it.

HILL 2 · INSTRUCTORAn instructor NEEDS TO spend critique time where their judgement changes the outcome, BECAUSE the same first-pass note thirty times leaves no time for learners who need them.

HILL 3 · ACADEMYcolorME NEEDS TO give every learner several feedback cycles without adding instructor hours, BECAUSE portfolio quality drives completion and referrals.

Problem statement

How might we give every learner specific, rubric-grounded feedback within minutes, in a way that strengthens their own eye, while instructors keep control?

03 Ideate & prioritize

Twelve ideas. The two easiest ones were unwise.

1Peer-crit circles with rubric

2Rubric self-check

3Common-mistakes video library

4AI auto-grader, score only

5AI coach: pinned notes + questions, reflect-first

6AI generates the “improved” poster

7Instructor voice notes

8Weekly live group crit

9Alumni mentor marketplace

10Cohort mistake digest

11Version compare v1 → v2

12Guarded coach chat

03 The concept

Crit: three rules that make fast feedback safe for beginners.

1

The learner judges first

No AI feedback until they state their intent and rate themselves on every rubric criterion. The gap between their rating and the crit is where learning happens.

2

The AI coaches, never designs

Pinned observation, one Socratic question and one principle per criterion. No generated images, rewritten copy or exact specs. “Do it for me” gets redirected.

3

The instructor decides what matters

Final work, low-confidence crits and self-vs-AI gaps ≥ 2 levels go to the instructor, who can agree, adjust or replace every AI judgement.

04 Prototype · live at colorme-crit.pages.dev

Specific, pinned, and honest about uncertainty.

Named the real issues: date competing with the title, three typefaces, low-value tagline, clip-art concept.
Each comment is pinned to the element.
AI level shown against the learner's own: colour 4 vs 2 → held for instructor.
≈ 10 s per crit, English or Vietnamese.Fact · test run

Synthetic learner and poster. Real model output.

04 Prototype · the loop

Plan → v2 → see what moved. The instructor sees what matters.

Learner: v1 → v2 with rubric deltas (“Developing → Proficient”). Average +1.2 levels on this synthetic example.

Instructor: confidence per criterion; agree / adjust / replace; note; timed review. Vietnamese crit shown.

04 How it works

A workflow, not an agent. AI assists · systems enforce · people decide.

AI assists

Reads the artwork (IDs 26, 22, 23), writes per-criterion feedback as JSON (27, 34), personalises to intent and language (20), answers and classifies chat (30, 32).

Excluded: ID 33, image generation.

Systems enforce

Reflect-first gate · confidence floor 0.6 · self-vs-AI gap ≥ 2 → hold · final → hold · prescriptive-output filter · daily limits · role-based access · audit trail.

People decide

Instructor: rubric, held crits, overrides, final grades. Learner: every design decision. Academic lead: go-live, prompt changes, pause.

01–02Experience & appResponsive web app: learner, instructor, pilot dashboard · Google sign-in
03–04Orchestration & AIclaim → critique (Gemini Flash vision, JSON schema, fallback model) → policy routing → release
05–07Data & infraInstructor-owned rubrics · artwork in R2 · crits, reviews, reactions in D1 · Cloudflare Pages Functions · CI/CD from GitHub

05 Agile · assumptions & MVP

Test the riskiest assumption first: does it agree with instructors?

#	Assumption	Cheapest test	Pass if
1	AI ≈ instructor judgement	Golden set: 60 past pieces, 2 instructors, blind	±1 level on ≥ 85%
2	Learners reflect first	Funnel: upload → reflection	≥ 80% complete
3	Faster feedback → more, better revisions	Treatment vs control classes	≥ 60% resubmit; +0.5 level
4	Instructors save time	Timed reviews + retro	−40% min/learner
5	Coach never does the work	Red-team prompts; audit 50 crits	0 prescriptive; 100% deflected
6	Learners consent	Consent screen, week 1	≥ 90% opt in

MVP: in (built)

One assignment (event poster), one 5-criterion rubric · reflect-first · pinned crit EN/VI · routing + instructor review · plan → v2 compare · guarded chat · pilot dashboard

Out, and why

AI grades (trust, fairness) · generated designs (does the work) · LMS integration (40 learners enrol by link) · other media (prove one first) · cohort digest (needs data)

05 Agile · delivery & learning loop

Every week: learn from disagreement, re-test, then release.

Cupcake · pilot6 weeks · 1 assignment · 2 classes

Birthday cake · foundations2 quarters · all assignments

Wedding cake · whole school12 months +

Experience

“A real crit on my poster the same evening.”

“Crits on every assignment; my instructor sees where the class is stuck.”

“From first poster to job-ready portfolio, I can explain my growth.”

Adds

Reflect-first crit, routing, review, versions, chat

Logo & social rubrics · cohort digest · rubric versioning · LMS sync · paid AI tier

UI/UX & motion crits · portfolio review before job applications · B2B training

Delivery: 2-week sprints · academic lead = product owner · 1 developer · both pilot instructors in every sprint review · evidence gates, not dates.

06 Pilot

2 classes vs 2 control classes · one poster · six weeks.

< 10 min

median feedback turnaround

baseline 2–4 days

≥ 60%

resubmit after feedback

vs control

+0.5

rubric level v1 → final, over control

blind-graded finals

−40%

instructor minutes per learner

timed reviews

≤ 20%

AI–instructor disagreement

overrides / criteria

Scale

All five hit · no incident · instructors want to continue

Adjust

1–2 miss, learning ≥ neutral → fix, one more cohort

Stop

Disagreement > 35% · learning below control · any privacy incident

06 Risks, Responsible AI & change

The real risks are human, so the controls are too.

Risk	Control
Confident, wrong feedback	Confidence floor + gap routing · visible confidence · “I disagree” · weekly calibration
Over-reliance	Reflect-first · questions not answers · plan in own words · no generated designs
Everyone designs “for the AI”	Level 4 rewards originality · instructors grade finals
Instructors feel replaced	They author the rubric, override everything, grade finals
Privacy, minors	Decree 13/2023 consent · guardian consent · paid tier, no training · no names in prompts
Silent model drift	Pinned model · golden-set re-run on any change · fallback · pause switch

NIST AI RMF

Govern named owners, change control · Map beginners, minors, IP · Measure scorecard by language & style · Manage routing, override, pause

Change management

Instructors co-write rubrics, see golden-set results first, 1-h onboarding, a champion. Learners “the AI can be wrong, disagree” from session 1. Academic lead owns a weekly 30-min calibration ritual.

07 Decision

Approve a 6-week controlled pilot: 2 instructors, ~40 learners, one assignment.

We need

A paid AI key with no training on data · a named academic owner · 2 instructors × ~6 h calibration · 2–3 dev weeks to add consent and a golden-set runner

It costs

≈ US$0.005–0.02 per critEstimate · under ~5,000 VND per learner per course · hosting in Cloudflare's free/entry tier

We'd stop if

Golden-set agreement < 75% · learners skip reflection · learning below control · instructors spend more time

Fast feedback is easy now. Feedback that builds the learner's own eye is the product.

Try it: colorme-crit.pages.dev · Code: github.com/hungkhi/colorme-crit · Appendix follows

×

Appendix A1 capabilities

AI suitability & capability map

ID	Capability	Use in Crit
26	Image captions	Describe elements & relationships
22	Identify objects	Locate elements → pins
23	Image → text	Read copy incl. Vietnamese diacritics
27	Image metadata	Level, confidence, box as JSON
34	Generate text	Observation, question, next step
20	Personalize	Intent, language, calibration, plan check
30 · 32	Q&A · intent	Coach chat; do-my-work detection
19	Moderate	Prescriptive-output filter
04 · 28	Patterns · summarize	Next: cohort digest
33	Generate images	Excluded by design

Is it really an AI problem?

Non-AI fixes (self-check, peer circles, videos) help and are included, but none gives specific feedback on this piece within minutes. Formative feedback tolerates labelled, contestable error; grades don't, so grades stay human.

Autonomy ladder

Formative crit = Recommend · Final work = Draft, human approves · Never acts alone. Climb only when errors are reversible, measurable and owned.

Appendix A2 spec

AI product specification

Item	Specification
Inputs	Artwork (≤ 8 MB, downscaled to 2,000 px) · brief · rubric (criteria × 4 descriptors) · intent (≥ 20 chars) · focus · self-rating 1–4 on every criterion · EN/VI · final flag · v2+: previous levels + plan
Outputs	Per criterion: level 1–4, confidence 0–1, observation, question, next step, box [ymin,xmin,ymax,xmax] on 0–1000 · 2 strengths · 1 priority · 1 reflection prompt · JSON-schema constrained
Routing	Hold if final · any confidence < 0.6 · \|self − AI\| ≥ 2 · missing criterion. Else release. Thresholds are pilot parameters.
Guardrails	No image generation in the product · prompt forbids fonts/hex/sizes/copy/layouts · post-filter rewrites prescriptive steps · chat intent classifier
Model	Gemini Flash vision + fallback model, 2 attempts each; prompt and schema are model-agnostic
Constraints	≤ 20 crits & 60 chat messages / learner / day · p90 < 60 s · < US$0.02 / crit
Acceptance	Golden set ≥ 60 double-marked pieces: ±1 level ≥ 85%, exact ≥ 60% · pins correct ≥ 80% · 0 prescriptive in 50-crit audit · 100% do-my-work deflected · EN/VI gap ≤ 5 pts

Appendix A3 evaluation

Multi-layer evaluation scorecard

Layer	Metric	Threshold
Business	Rubric improvement v1 → final vs control · resubmission · completion	≥ +0.5 · ≥ 60% · no drop
UX	Reflection completion · comments rated helpful · would use again	≥ 80% · ≥ 70% · ≥ 75%
Output quality	Agreement with instructor consensus (±1 / exact) · pin accuracy	≥ 85% / 60% · ≥ 80%
Safety	Prescriptive outputs in audit · do-my-work deflection	0 · 100%
Fairness	Agreement gap EN vs VI; hand-drawn vs digital; Vietnamese vs Latin type	≤ 5 pts
Robustness	Low-res, blank, non-design, text-heavy uploads	Low confidence → held
Latency · cost	p90 crit time · cost per crit (stored per crit)	< 60 s · < US$0.02

Appendix A4 architecture

Eight layers · two flows · build/buy/reuse

01ChannelsResponsive web app; later embedded in colorme.vn homework pages
02Business appsAssignments, submissions, versions, reviews, responses; later LMS sync
03OrchestrationDeterministic workflow: claim → critique → route → release
04AI servicesGemini Flash vision + JSON schema; fallback; policy module
05Data & knowledgeInstructor-owned rubrics & briefs; R2 artwork; D1 records
06IntegrationGoogle OAuth (PKCE) · Gemini API · GitHub Actions → Cloudflare
07InfrastructureCloudflare Pages + Functions, D1, R2
08Security & governanceHashed HttpOnly sessions · roles · per-learner file isolation · rate limits · audit trail · pause switch

Serve users

Upload & reflect → authorize → critique → route → instructor review → release & plan

Keep current

Rubric / prompt / model change → version → re-run golden set → academic lead approves → release

Build workflow & experience · Buy the model as an API · Reuse Google identity, Cloudflare · Crosses the boundary: artwork + intent to Gemini → paid tier + consent in production.

Appendix A5 backlog & change

Prioritized backlog · what changes

Pri.	Story	Status
Must	Rate myself before feedback	Done
Must	Comments pinned & tied to criteria	Done
Must	Uncertain/final crits come to instructor; override any level	Done
Must	Plan → v2 → see what moved	Done
Must	Consent (incl. guardians) before first upload	Sprint 1
Must	Golden-set runner gates every change	Sprint 1
Should	Rubric versioning · queue filters	Sprint 2
Should	Weekly cohort digest	Sprint 3
Won't	AI grades · AI-generated designs	By design

Dimension	With Crit
Experience	Private, pinned crit in minutes; questions, not verdicts; visible progress
Process	Reflect → crit → plan → v2 → instructor-graded final
People	Instructors author rubrics & review exceptions; academic lead owns quality
Data	Every version, rating, override and reaction structured and measurable
Operations	Weekly calibration, dashboards, pause switch, manual fallback

Appendix A6 dashboard & cost

Live dashboard computed from the database. Synthetic demo data shown, n = 3 submissions.

Cost range Estimate

Item	Pilot	Scale / yr
AI inference	~400 crits, under US$10	$0.005–0.02 per crit
Hosting	Free tier	~US$5–50 / mo
Development	2–3 person-weeks	0.5 FTE
Instructor time	~12 h calibration	~2 h / week

Optimise cost per successful outcome (a learner +1 level on a criterion), not cost per API call. Token prices to be re-verified before scale.

Appendix A7 evidence & AI use

Evidence, limitations and AI-use disclosure

Evidence tags

Fact verified (colorme.vn, prototype runs) · Estimate calculated · Assumption founder's view · Hypothesis pilot tests it. No interviews or results are presented as real. All demo accounts, posters and metrics are synthetic.

Limitations

No primary research yet (week 0 adds 5 learner + 2 instructor interviews) · model confidence is self-reported, so the 0.6 floor must be calibrated · prototype uses a free-tier AI key with synthetic data only.

AI-use disclosure (Level 4, syllabus §7)

Claude Code (Claude Opus 5): read the course materials on Canvas to align structure and vocabulary; drafted the report and this deck; built, tested and deployed the prototype on my direction.
Gemini Flash: the AI inside the product (critique, coach chat).
HTML + headless Chrome: synthetic demo posters.
Checks: type checks, 7 unit tests (routing, overrides, guardrail), live end-to-end runs, CI on every push; every colorME number tagged by evidence type.
I remain accountable for every claim and decision. Details in report §17.

Appendix A8 why portfolios

The artefact every course builds towards

2025–26 portfolios published on Behance by Vietnamese designers. A colorME learner is measured against this when they apply for work — which is why feedback cycles per piece, not hours taught, is the number that matters.

Covers reproduced for academic illustration; all rights remain with their authors (full credits in the report, §18). None of these designers is claimed as a colorME learner.