The REMEMBR app icon: a cream rounded square with a mark of concentric orange arcs resembling a fingerprint.

MRP Thesis Defense

REMEMBR

Building an accessible journaling application without a programming background.

Raafi Islam Master of Digital Media · Toronto Metropolitan University

Agenda

How this defense runs

  1. The problem

    Why I built this. A problem I kept running into, and what the apps we already have could not do about it.

  2. Methodology and Positionality

    What evidence I used, and what studying my own practice can and cannot prove.

  3. REMEMBR

    The app I built, and the accessibility thinking behind each decision in it.

  4. Findings

    The four skills this took, the four ways the AI kept failing, and the three habits I built to work around them.

  5. Discussion

    My answers to both research questions, what this adds, and where it goes next. Then your questions.

Dhaka, 2018

I recognized the face.I couldn't place the name.

It was short and it was awkward. I made an excuse and left. What stayed with me was not the moment. It was realizing I had no way back to that person. I had no record of how I knew them, who introduced us, or what we had talked about.

A Recurring Pattern

It did not happen once.

It kept happening, until I started avoiding situations where it might come up again.

Epilepsy

Epilepsy affects how memories get stored in the first place. It is my own diagnosis, and the reason this kept happening to me.

Monthly support sessions

I heard the same frustration from other people, month after month. The tools we have are not built for memory you cannot rely on.

The Gap

Journaling apps organize by date. Contact apps organize by name. Associative memory does neither.

Standard journaling applications, Day One, Journal, Diary, organize content by date. Standard contact applications organize by name. Neither model reflects how associative memory actually works.

People are more reliably retrievable through their relational context and associated events than through a name or a face alone (Yovel and Duchaine, 2006). REMEMBR grew out of that gap: entries organized around people rather than dates, surfaced through any connected context. A shared location. A mutual contact. An event.

Research Questions

Two questions drive this MRP

  1. RQ1

    How can a non-programmer use generative AI to design and build a functional software application, and what specific competencies does that process require?

  2. RQ2

    What are the practical limits, recurring failure modes, and design trade-offs of AI as the primary development tool, and how were they navigated?

Thesis

Generative AI has created a genuinely new pathway for non-programmers to build production-grade software. It is not frictionless.

It demands a distinct skill set, rooted in design, editorial, and project-management practice, not programming. The skills are real. So are the limits.

Methodology and Positionality

How I studied this. What evidence I used, and what research into your own practice can and cannot prove.

Evidence Base

Two kinds of primary evidence

AI conversation transcripts

My ChatGPT conversations from October 2025 to April 2026, and nine Claude Code sessions from April to June 2026. Together they record every instruction I gave, what the AI sent back, and how I corrected it.

Versioned application files

Two separate version histories. Fifteen named versions from the early build, each saved as its own file, and a second run of versions from the first five days with Claude Code.

Positionality

Designer, director, and subject

I did not begin with a research framework. The process was driven by personal experience and iterative making: trying things, observing what happened. This resembles what design researchers call research through practice (Frayling, 1993): knowledge generated by the act of making, not by observing from outside it.

What this makes possible

I know why every decision was made, which nobody watching from outside could reconstruct. At every step I knew exactly how far the output was from what I wanted.

What it cannot guarantee

I cannot be fully objective about my own experience. I handle that by tying each claim to a dated transcript rather than to memory.

Limitations

What this study cannot claim

A single case

One case cannot tell you what is generally true of non-programmers building with AI.

No external verification

The dated transcripts back up what I claim, but I am the one reading them.

Theory found after the fact

I noticed the link to Actor-Network Theory afterwards. I did not set out to test it.

What it can produce: a detailed, internally consistent account of one instance of this process, precise enough to generate propositions later research can test.

REMEMBR

The app this project produced, and the accessibility thinking behind every decision in it.

The Artifact

A people-first journal, not a dated diary

You build a profile for each person in your life, linked to places, events, and people you both know. Remember any one of those things, and you should be able to find the person.

Built for

Epilepsy · Prosopagnosia · Traumatic brain injury · ADHD

8 months
From October 2025
13,000
Lines, in a single HTML file
0
Lines of production code written by hand
REMEMBR home screen on a dark background. A greeting reads Good morning, Raafi. Below it a Your People row shows circular portraits of Adam, Olivia, Daniel and Sara, then prompt cards reading Who did you meet today? and Siobhan's birthday in 2 days.
Figure 9. Interface. Generated with Claude (Anthropic, 2026).

Design Principle

Accessibility is the baseline, not a feature

Governed throughout by WCAG 2.1 AA, Ontario's AODA, and W3C COGA guidance for cognitive accessibility. Every setting in this panel is user-controlled, not auto-detected.

Perceivable
Atkinson Hyperlegible type. Color went through five rounds before landing on a sky-teal palette that passes WCAG 2.1 AA. Anything saved can be read out loud.
Operable
You can dictate into any text field, not just a special voice mode, and it is tuned for a wide range of accents. Tap targets are larger than 44 by 44 CSS pixels. Motion can be turned down, and there is no high-frequency strobing.
Understandable
Easy Mode replaces every multi-field form with one plain question at a time.
Robust
ARIA labels, full keyboard control, and screen reader support.
REMEMBR's Accessibility settings panel. Labelled toggles for Easy Mode, High Contrast, Reduce Motion and Haptic Feedback, a Text Size control set to 100 percent, and a Voice and dictation language selector.
Figure 3. Accessibility Features. Generated with Claude (Anthropic, 2026).

Signature Feature

The Connection Web

A graph where people, places, memories, and events all carry the same weight. Start anywhere, from a mutual friend or an event, and follow the links out to the person you are trying to reach.

Structurally resonant with Actor-Network Theory (Latour, 2005): a connection recognized only in retrospect, after the interface was already built from lived experience rather than from theory.

The Connection Web in REMEMBR, titled Adam's Web with four connections. Adam sits at the centre of a dark canvas with thin teal lines radiating out to Daniel above, Olivia to the right, Sofia below and Jesse to the left, with a list of the four connected people beneath the graph.
Figure 4. Connection Web. Generated with Claude (Anthropic, 2026).

Any point is an entrance

Signature Features

Two more ways in

Associative Search

One search box looks through everything at once, so you never have to pick a category first. That matches how episodic memory is actually retrieved, through linked cues (Tulving and Thomson, 1973).

REMEMBR search results for the single word coffee. Filter chips read All, Colleague, Friend, Family and Acquaintance. Two matching people, Priya Nair and Sara Williams, appear as cards with photos, how they were met and a relationship label. Below, a heading reads 14 memories found, followed by entries including Full circle: Priya meets Marcus at Balzacs Coffee with Daniel and Priya, Coffee catch-up with Adam, and Sara mailed a watercolour.
Search, current build (cf. Figure 5). Generated with Claude (Anthropic, 2026).

Easy Mode

Every multi-field form turns into one question at a time, right across the app. It is grounded in cognitive load theory (Sweller, 1988).

REMEMBR in Easy Mode. The home screen asks What would you like to do? and offers four large, plainly labelled choices: Add a person, Add a memory, See my people, and My memories.
Figure 2. Easy Mode. Generated with Claude (Anthropic, 2026).

Signature Features

Designed to remove friction, not add polish

Avatar Creator

Not everyone has a photo of every person they want to remember. I set Snapchat's Bitmoji as the bar to hit, and rejected three rounds before the fourth was good enough.

REMEMBR's avatar creator. A large illustrated face preview sits above a Randomise button, a row of eight selectable skin tone swatches, and a grid of five face shape options with the last one selected.
Figure 6. Avatar creation. Generated with Claude (Anthropic, 2026).

Social Card Generator

A card you can share that tells someone what they mean to you. It reads: "I REMEMBR YOU."

REMEMBR's Share your memory screen. A generated card for Adam Chen, labelled Colleague, reads We met at DG8011 class on Thu, Sep 5, 2024. Sat next to each other on the first day, above the tracked words I REMEMBR YOU.
Figure 8. Social Card. Generated with Claude (Anthropic, 2026).

Findings

Eight months of working across several AI tools and models, running trials, exploring options, and fixing errors. Here is what came out of it.

Finding 1 of 3

Four competencies, not one line of code

  1. Specification

    Turning what you want into words precise enough for the AI to build it correctly. You do this constantly, not once at the start.

  2. Evaluation

    Judging whether the result is good enough without reading the code. You go on how it behaves, how it compares to other products, and what testing shows.

  3. Direction

    Giving a correction specific enough to actually fix things. Naming a product to match works far better than saying "make it better."

  4. Continuity Management

    Holding one clear picture of the product across dozens of separate sessions, with tools that remember nothing in between.

Finding 2 of 3

Four failure modes, recurring by structure

  1. Silent Feature Regression

    Features I had already signed off would disappear between sessions, with no warning. The AI only saw what was in front of it, not a running list of what already existed.

  2. Persistent Hallucination

    The same wrong facts kept coming back. It repeatedly said the app stored everything on the device, when REMEMBR actually syncs to the cloud.

  3. Token Limit Constraints

    Useful working time dropped to 15 or 20 minutes before the context ran out. Waiting for the limit to reset could take up to five hours.

  4. Code Token Exhaustion

    At 13,000 lines, simply loading the file into a new session used up a large part of the context before any work began.

Finding 3 of 3

Three practices developed in response

  1. Structured Context Documents

    A dense briefing pasted in at the start of every session, because a chat tool opens each conversation blank. Claude Code made this easier later on. It keeps project notes in files on the machine, so the context carries over instead of being pasted again.

  2. Standing Rules

    Every mistake the AI kept repeating became a fixed rule I carried into every later session. I never had to explain it twice.

  3. Forced Review and Planning Cycles

    The forced waiting turned into useful time. I used it to test what was live, find bugs, and plan the next features on paper.

A Question This Raises

Who owns this, and is the data safe?

REMEMBR holds private memories about real people. Who owns the code, and who can see the data, were requirements from day one, not questions I left to the end.

Code ownership
The law here is still settling. Copyright needs a human author, but work where a human directs the AI through creative choices can still qualify (U.S. Copyright Office, 2023). Anthropic does not claim to own what it generates (Anthropic, 2024).
Training data
I kept the training opt-out on the whole way through, so the sessions where REMEMBR was built were excluded from model training (Anthropic, 2024).
Platform choice
I left Rork partly because keeping project data private needed a paid subscription. REMEMBR runs on infrastructure I own and can check myself.
Illustration of three cartoon figures in a tug of war over a large phone showing an app icon. One wears a ChatGPT t-shirt, one a Claude hoodie, one a Google sweater, with loose papers and a USB cable scattered on the floor.
Everyone wants a piece of the app you build.

Discussion

Answering both research questions

  1. RQ1

    How can a non-programmer use generative AI to design and build a functional software application, and what specific competencies does that process require?

    A non-programmer can build production-grade software by developing four specific competencies: design and editorial skills, not programming ones, already held by a population far larger than the one that currently builds software.

  2. RQ2

    What are the practical limits, recurring failure modes, and design trade-offs of AI as the primary development tool, and how were they navigated?

    The limits cluster into four failure modes, all expressions of one root cause: a stateless tool operating on an incomplete picture of what it has already built. These are manageable through documented, repeatable practice.

Contribution and Next Steps

What this case adds, and where it goes

To the literature

Four skills, defined clearly enough that other people can test them. And a context-document practice that answers a problem the existing research does not address: working across many sessions with a tool that remembers nothing.

What comes next

Keep building the same way. Test it with users other than me. Move from closed testing to a full release.

Ko et al. (2011) named the persistent gap between the people who need software and the people who can build it. REMEMBR is a direct instance of that gap narrowing, not because the technology got smaller, but because the skills it demands are already held by a much larger population than the one that currently builds software.

I REMEMBR YOU

Thank you.

Questions and discussion

Raafi Islam · REMEMBR · Master of Digital Media, Toronto Metropolitan University

References

Works cited in this presentation

  1. Anthropic. (2024). Privacy policy. https://www.anthropic.com/legal/consumer-terms
  2. Anthropic. (2026). Claude [Large language model]. https://claude.ai
  3. Braille Institute of America. (2019). Atkinson Hyperlegible font. https://www.brailleinstitute.org/freefont
  4. Frayling, C. (1993). Research in art and design. Royal College of Art Research Papers, 1(1), 1-5.
  5. Government of Ontario. (2005). Accessibility for Ontarians with Disabilities Act, 2005, S.O. 2005, c. 11.
  6. Ko, A. J., Abraham, R., Beckwith, L., Blackwell, A., Burnett, M., Erwig, M., Scaffidi, C., Lawrance, J., Lieberman, H., Myers, B., Rosson, M. B., Rothermel, G., Shaw, M., & Wiedenbeck, S. (2011). The state of the art in end-user software engineering. ACM Computing Surveys, 43(3), 21:1-21:44.
  7. Latour, B. (2005). Reassembling the social: An introduction to actor-network-theory. Oxford University Press.
  8. Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257-285.
  9. Tulving, E., & Thomson, D. M. (1973). Encoding specificity and retrieval processes in episodic memory. Psychological Review, 80(5), 352-373.
  10. U.S. Copyright Office. (2023). Copyright and artificial intelligence. https://www.copyright.gov/ai/
  11. W3C COGA Task Force. (2021). Making content usable for people with cognitive and learning disabilities. World Wide Web Consortium.
  12. W3C Web Accessibility Initiative. (2018). Web Content Accessibility Guidelines (WCAG) 2.1. World Wide Web Consortium.
  13. Yovel, G., & Duchaine, B. (2006). Specialized face perception mechanisms extract both part and spacing information: Evidence from developmental prosopagnosia. Journal of Cognitive Neuroscience, 18(4), 580-593.

Note. The REMEMBR screens shown here are figures from the MRP and keep the paper's figure numbers, except the search screen on slide 17, which is a later capture of the same feature. All show software generated with Claude (Anthropic, 2026).

REMEMBRMRP Defense

Press → to continue