Curb Cut

Curb Cut lets any worker ask for a workplace accommodation in their own words, on the web with no login, by text or a call from a basic phone, by email, in a Slack DM, or through their own assistant, without ever disclosing a diagnosis. An Agentforce agent holds the conversation; deterministic Apex decides what is stored and sent; a person on the access desk answers. First place, Builder Track, Agentforce for Good Hackathon, Dreamforce 2026.

git clone https://github.com/parthsevak2/curb-cut
cd curb-cut
python3 tests/invariants.py     # 521 structural checks, offline, about a second

Feel it before you read anything: type one sentence at parthsevak2.github.io/curb-cut/try.html and watch the product's own ranking answer, entirely in your browser. The landing page with the films is parthsevak2.github.io/curb-cut, and a print-ready break-room poster is poster.html.

The story behind it, ten minutes: https://parthsevak2.github.io/curb-cut/story.html

To install it into your own Salesforce org, follow DEPLOY.md. It needs Agentforce enabled; text and voice also need your own Twilio number and a small relay (channels/), and Slack needs the app created in your workspace from the manifest. Four Custom Labels carry your deployment's number, site URL and support address.

The process for requesting a disability accommodation is itself an accessibility barrier.

To get help you must first navigate a system designed by and for people without your disability. Find the policy. Learn its vocabulary. Hold a corporate login. Complete a form. And disclose something you may be afraid to disclose, to a stranger, in writing, in a language that may not be your first.

Cost usually isn't the barrier.

The film, 4 minutes 59 seconds, with every channel shown live: https://youtu.be/yleHLiwWRKA. A 26-second recording of trying the ask page: https://drive.google.com/file/d/13NApu8rq3MpJjuGG2yzKuJ2P02tKWwGx/view. A three-minute cut: https://drive.google.com/file/d/1Loqum6Kbk1noQazJDKrrVEnHzoU3biMD/view

Curb Cut is an Agentforce agent whose principal is the worker, not the employer.

The hackathon deployment below is kept as a record of what judges tested. The demo org and its number retire with the hackathon; your install has its own addresses on every row.

The console is the half of this that nobody demos. Every promise the assistant makes about a person reaching a human being is worthless unless a human being has somewhere to stand. Human_Handoff__c was creating records that said "a person is picking this up" with no surface for that person.

Nothing in the console shows a diagnosis, because there is no field for one. An operator who is asked what a person has can only answer that nobody knows and nobody can find out.

1. The agent's principal is the worker. Nothing leaves that the person did not personally approve.
2. No channel is the real one. Text, voice, email and web are equal front doors.
3. Ask before you disclose. A real answer without identifying yourself or naming anything about your body.
4. Ask once, not forever. Standing preferences travel ahead of you.

• No automated sign language. ASL is a complete language; recognition is unsolved. Signed video routes to a human, with immediate text acknowledgement so nobody waits in silence.
• No diagnosis field. It isn't encrypted or restricted. It's absent. CI fails the build if anyone adds one.
• No phone fallback, ever. Reachable_By__c accepts a message channel or email, and the code refuses a telephone number.
• No inference. Nothing is profiled or predicted. Every signal is one the person set.

Nothing below is asserted from memory. Each number is the output of the command beside it, and tests/sync_counts.py rewrites the published counts from the invariant suite's own output so they cannot drift.

147 Apex tests                 sf apex run test -o curbcut -l RunLocalTests
554 structural invariants      python3 tests/invariants.py            ~1s, no org
517 accessibility checks       python3 tests/a11y_audit.py            against the live pages
122 Sa11y checks               npm run test:a11y                      Salesforce's own axe-core matcher
 34 accessibility-tree checks  node tests/ax_tree_audit.mjs           what a screen reader is handed
 34 controls by keyboard       node tests/keyboard_walk.mjs           real Tab presses; set CURB_CUT_SITE
 21 responsible-AI checks      python3 tests/rai_self_check.py        against Salesforce's five guidelines
 28 contrast checks            python3 tests/contrast_audit.py        both themes, from the tokens
 16 reading-level checks       python3 tests/reading_level_audit.py
 76 link and copy checks       python3 tests/link_and_copy_audit.py
23/23 adversarial assertions   node tests/headless_agent_api.mjs && python3 tests/score_adversarial.py

Two of those suites break their own checks on purpose and assert each one goes red (--selftest). A suite that cannot fail proves nothing about the code it passes. The adversarial suite once scored 21 of 23, and both failures were published with their transcripts rather than averaged away; the agent's refusal wording and the scorer have since converged, and the suite passes whole. The habit stands: a miss gets reported, never smoothed.

Start with docs/ARCHITECTURE.md, then CRITIQUE.md, which is where the honest list of what is still weak lives.

Built by Parth Sevak. Welland, Ontario, Canada. Agentforce for Good, Dreamforce 2026 · Builder Track · Abilityforce.

Apache 2.0. The accommodation library ships in the repo with a source citation on every row. Use it, fork it, run it in your own org.