UX CASE STUDY · GOVTECH · Q1 2026

A court portal built to answer in under 2 minutes with a clerk verifying every AI response.

A proof of concept portal that gives self-represented litigants clerk-verified answers, remote filing, and affordable AI transcripts. Built from four years of watching where the court process locks people out, designed against a 72 hour legacy baseline.

< 2 mindesign target for a clerk-verified answer, against a 72 hour help desk
~40%of filings rejected under the legacy process
$9.25per page for a priority transcript, the cost the flat daily rate replaces
21court system actors, mapped from four years inside In-Court Technology
Role
Lead UX & Dev
Timeline
Q1 2026
Stack
Figma · Figma Slides · Gemini 3.1 Pro
Constraint
Proof of Concept
WALKTHROUGH003

See It Run

Forty seconds, no sound. A question, the clerk queue, a filing and a transcript request.

VID. 01Real captures of the clickable A2J prototype, which runs on canned data: a plain-language question with its answer pending clerk verification, the clerk queue sorted by risk, the 24/7 e-filing form, and the flat-rate transcript request.
PROBLEM → SOLUTION004

The Problem vs. The Solution

Courts run on dense legal language, in-person filing, and expensive transcripts. After four years at the Supreme Court of BC fielding the public's questions, I knew exactly where people got stuck, so I designed the portal I wished the registry desk could offer.

BEFORE004.A

The legacy process locks people out

Dense legalese, mandatory travel, and per-page transcript pricing combine to make the court system hardest on the people with the least slack.

P-01
Legalese overloadDense legal language creates a barrier to entry, driving submission error rates near 40% and sending rejected filings back through the queue.
P-02
Physical barriersMandatory in-person filing penalizes rural, disabled, and hourly-wage individuals who cannot spare unpaid time to stand in a courthouse queue.
P-03
Transcripts priced out of reachCourt reporters charge up to $9.25 CAD per page for priority requests, pricing low-income litigants out of their own records.
AFTER004.B

Clerk-verified AI, remote by default

An AI plain-language engine with a human verification layer, wrapped in a portal that works from any device at any time.

S-01
AI plain-language engineAI summarizes complex codes in plain English and guides users through forms. Clerks cross-reference every output against legal citations before users see it.
S-02
100% remote architectureFiling, responses, and case file access from any device at any time. No travel, no queues, 24/7 cloud access mapped to user profiles.
S-03
Flat-rate AI transcriptsSame-day searchable AI transcripts at a flat rate per day requested, replacing per-page pricing and unskimmable raw audio.
OLD vs NEW WORKFLOW005

The User Flow, Before and After

Two journeys to the same filing. One is defined by travel, queues, and rejection loops; the other by a plain-language path a litigant never has to leave home for.

OLD WORKFLOW · 3-7 DAYS005.A
  1. 01Identify the legal need, no guidance offered
  2. 02Research the requirements in dense legalese
  3. 03Travel to the courthouse, 45 to 90 minutes each way
  4. 04Queue at the help desk, 45 minute average wait
  5. 05Complete dense paper forms by hand
  6. 06Submit at the counter, no digital receipt
  7. 07Await a response, 24 to 72 hours
  8. 08Rejected filing forces a return trip, start over
NEW WORKFLOW · UNDER 5 MINUTES005.B
  1. 01Open the portal from any device, zero travel
  2. 02AI explains the step in plain language, instantly
  3. 03Guided smart form catches errors inline
  4. 04Submit digitally, instant timestamped receipt
  5. 05A clerk verifies the AI answer in under 2 minutes
  6. 06Track the case file 24/7 from the dashboard
3-7 DAYS → UNDER 5 MINUTES
KEY DECISIONS006

Key Design Decisions

Three calls that shaped the concept, each anchored to a constraint I saw firsthand at the court.

DEC-01 Human verification on every AI answer A wrong legal answer can cost someone their case. No court can publish unverified AI output, so the verification layer is what makes the AI usable at all. What happened

AI drafts the plain-language summary, but clerks cross-reference it against legal citations and users only see clerk-approved steps.

DEC-02 Flat-rate transcript pricing Per-page pricing scales with how much justice you can afford, and it was pricing low-income litigants out of proper representation. Financial accessibility is the ultimate usability feature in a government product. What happened

Same-day AI transcripts billed at a flat rate per day requested instead of up to $9.25 per page.

DEC-03 Privacy as a hard requirement Court files hold sensitive personal data, and at the court I watched policy review kill projects that raised security questions. Designing for privacy up front is what keeps a concept like this shippable. What happened

Data handling was scoped conservatively from the start, carrying over the security-first approach of every tool I built at the court.

THE SYSTEM MAP007

The court system, mapped from the inside

Before the portal had a single screen, it had this map. Twenty-one actors in the BC court system, drawn from four years inside In-Court Technology, with every relationship labelled the way it works in practice rather than the way the org chart says it should.

REGISTRY The front door that cannot advise

Everyone files, asks, and queues at the counter, and the counter is procedurally barred from giving the one thing a self-represented litigant needs most.

ICT The hidden broker

Audio requests, vendor screening, and the weekly 525 bail-review video list all route through my old team. The middle of this map is where I worked.

SRL vs COUNSEL Four channels against one

A firm files through agents, orders through vendors, and researches with paid tools. The self-represented litigant gets the counter that cannot advise them.

COURT ACTORS · INTERACTIVE007.AOpen the interactive map

Why the board is silent at rest. Every relationship label stays hidden until you hover or tab to an actor, so the first read is structure, not noise. Interest lights up exactly one actor's world: its connections, its edge labels, its notes from the operator's seat. That quiet-at-rest pattern is how I keep dense operational diagrams legible. The reader chooses the detail; the map never shouts it.

This is Map 1 of a four-map service blueprint behind the portal. Information flows, the case journey through a self-represented litigant's eyes, and a gap map that feeds the roadmap all follow the same pattern.

LIVE PROTOTYPE008

Walk through the A2J Portal

The conceptual portal, running. Move through the plain-language guide, the smart form, and the clerk-verified answer, from any device. This is a prototype, not a production system.

a2j.kashb.ca Open in new tab ↗
REFLECTION009
Designing from the registry desk

Most govtech concepts start from a technology looking for a problem. This one started from four years of answering the public's questions at the registry desk, where I watched the same people hit the same walls every week. The AI is not the interesting part. The clerk verification layer is, because it is the piece a real court could actually adopt. The honest next step is a pilot with a live registry to turn these design targets into measured outcomes.

Under the hood

The access model, plus the standards and process behind the portal. Closed until you want it.

Who can see whatThe access model, eight kinds of record against five kinds of person.1 min read
ACCESS MODEL010

Who Can See What

A court file isn't one document with one audience. Eight kinds of record, five kinds of person, and almost none of those pairs are a simple yes or no. This is the access model I designed the portal around.

PERMISSION MATRIX010.A
Access by role and record type, file BC-2026-CV-0892
Record Public Litigant Counsel Clerk Judiciary
Case metadata and hearing dates ● ● ● ● ●
Their own filed documents ○ ● ● ● ●
Opposing party filings ○ ● ● ● ●
Clerk log notes × ◐ ◐ ● ●
Court audio recording ○ ○ ○ ● ●
Transcript orders × ● ● ● ●
Sealed or restricted documents × ◐ ◐ ◐ ●
Publication-ban content × ◐ ● ● ●
Full On request None Redacted
Public

Hearing dates and the style of cause, nothing more without asking. Filed documents come through a records request. Audio needs a listening session booked through the registry, which is a real counter transaction, not a download.

Litigant, self-represented

Their own filings and the other side's, in full. Clerk log notes stay internal. Audio still goes through the registry, same as anyone.

Counsel of record

Everything a litigant sees, plus publication-ban material they need to argue the case. Audio on an undertaking.

Registry clerk

The working view. Log notes, transcript orders, the audit trail. Sealed content stays closed to them too, because handling a file isn't the same as being entitled to read it.

Judiciary

Full file, including sealed material. The only role with no redaction.

This is the portal's designed access model, built from four years of handling these requests at the registry counter. It's not a statement of BC court records policy, and a real deployment would get scoped with the court's records office.

WCAG 2.2 AA and the GC Design SystemWhat each standard changed in the design.2 min read
STANDARDS ALIGNMENT011

What This Was Built Against

Government work has a floor. WCAG 2.2 AA and the GC Design System are it, and each one changed the design in ways worth naming.

WCAG 2.2 AA011.A

The six that bite at AA

WCAG 2.2 landed in 2023. Most accessibility write-ups still quote 2.1. These are the new ones, and what each cost this design.

3.2.6A
Consistent Help

The route to a human clerk sits in the same spot on every screen. It never moves to make room for anything.

3.3.7A
Redundant Entry

Case number, party names, and address carry through the filing flow. Nobody types their case number twice.

2.4.11AA
Focus Not Obscured

The progress bar and the help bar are both sticky, so both had to be checked against every focused field. A sticky header that covers the field you just tabbed into fails this.

2.5.7AA
Dragging Movements

Exhibit ordering has up and down buttons. Drag works, but nothing needs it.

2.5.8AA
Target Size (Minimum)

The floor is 24 by 24 CSS pixels. Everything here is 44 by 44, because the people using this are often stressed and on a phone.

3.3.8AA
Accessible Authentication

Sign-in is a case number and an emailed link. No puzzle, no memorized secret, nothing that tests your memory as a condition of filing.

One more, carried over from 2.1. The matrix above codes every cell with a glyph and a colour, not a colour alone. That's 1.4.1, and colour alone would have failed it.
GC DESIGN SYSTEM011.B

Federal patterns, provincial court

The GC Design System is the Canadian Digital Service's component library for federal services. Bilingual by default, Angular, React and Vue, with a paired Figma library.

JUR
Why a federal system for a BC concept

The Supreme Court of British Columbia is provincial. GC Design System is federal, so none of it binds here. I designed against it anyway. A pattern that already cleared federal accessibility and official-languages review beats one I invented, and a BC concept that speaks the federal system's language is one that can move.

LANG
Bilingual authoring, not translation

The part worth stealing is how the bilingual content gets authored. Labels, error text, and help copy are written as English and French pairs from the start, not translated at the end. That order does more than it sounds like it does. Writing both at once kills the clever English phrasing that has no French equivalent, and the English comes out plainer for it.

STATE
The toggle holds mid-filing

Switching languages partway through a filing doesn't dump your progress. Sounds obvious. It is the single most common place a bilingual form breaks.

FORM
One question per screen

Borrowed straight from the GC Forms pattern. Long court forms are where self-represented filers give up, and the error rate at the counter said so before any research did.

Reference · GC Design System · cds-snc
Design processFrom the legacy journey to a high-fidelity prototype.1 min read
DESIGN PROCESS012

Design Process

From mapping the legacy journey to a high-fidelity prototype, with clerk verification designed into the AI logic at every step.

01
DISCOVER
  • ▸Drew on four years fielding public questions at the court registry
  • ▸Documented the legacy friction: 40% error rates, 72 hour responses
  • ▸Mapped the before and after journeys in a full user flow comparison
02
DEFINE
  • ▸Set clerk verification as a hard requirement for every AI answer
  • ▸Scoped the portal to filing, responses, case files, and transcripts
  • ▸Committed to plain language over legalese on every screen
03
DESIGN
  • ▸Built the end-to-end UX strategy deck in Figma Slides
  • ▸Designed the human-verified AI logic: summarize, verify, approve
  • ▸Trust badges show users only clerk-approved procedural steps
04
DELIVER
  • ▸Shipped a high-fidelity interactive prototype of the portal
  • ▸Published the full case study deck and user flow comparison
  • ▸Labeled the work honestly as a conceptual prototype
The decksThe UX strategy deck and the core case deck, live from Figma Slides.1 min read
UX ARTIFACTS013

The Decks

Two presentations carried this project. The end-to-end UX strategy, and the core case deck that walks through the clerk-verified AI logic. Both are built in Figma Slides and embedded live below.

a2juxprocess.kashb.ca Open in new tab ↗

UX process deck. Discovery and diagramming through to the exact interaction logic, in the order the work actually happened.

a2jcase.kashb.ca Open in new tab ↗

Core case deck. The human-verified AI logic in three moves. AI summarizes the code in plain English, a clerk cross-references it against the citations, and users only ever see clerk-approved steps.

Available for Work

Let's build something impactful.

I'm currently open to new opportunities in UX/UI Design, Product Design, and GovTech transformation roles.