Research and Policy Analyst · Papers and presentations

Lightning RoD: AI Hackathon presentation

With a teammate, I presented Lightning RoD, a prototype that drafts the record of an Employment Insurance decision, at the department’s 2026 hackathon. This three-page summary, written for this site, covers parts of that presentation.

Public summary · three pages, written for this siteHow it works → the officer stays in charge → how it was built
The officer decides. The AI fills only the blanks.A record of decision drawn as an outline, with no words from any case. Approved wording: the program: The legal wording, the decision and the references, word for word. The AI cannot change them. The blanks: the AI: Written only from the officer’s notes, in plain language. The officer can edit every word. Gaps: the officer: Where the notes don’t support a reason, the AI leaves a gap instead of guessing. Approval waits until the officer fills it. In the outline the approved wording is shown as grey bars, the blanks the AI writes are shown in green, and one gap is marked “[Officer to complete]”. HOW IT WORKSThe officer decides. The AI fills only the blanks. A RECORD OF DECISION, AS AN OUTLINE THE FACTS THE DECISION THE REASONS [Officer to complete] THE REFERENCES Drawn without data: no case is shown. Approved wording: the program The legal wording, the decision and the references,word for word. The AI cannot change them. The blanks: the AI Written only from the officer’s notes, in plainlanguage. The officer can edit every word. Gaps: the officer Where the notes don’t support a reason, the AI leavesa gap instead of guessing. Approval waits until theofficer fills it. Lightning RoD: AI Hackathon presentation · public summaryThomas Eapen · 1 / 3

How it works

The officer decides. The AI fills only the blanks.

Approved wording: the program

The legal wording, the decision and the references, word for word. The AI cannot change them.

The blanks: the AI

Written only from the officer’s notes, in plain language. The officer can edit every word.

Gaps: the officer

Where the notes don’t support a reason, the AI leaves a gap instead of guessing. Approval waits until the officer fills it.

On a wider screen the record is drawn as an outline, without data: no case is shown.

The officer stays in chargeFairness: Both sides go on the record. The officer weighs them. Bias: Only facts from the notes. No opinions, no guesses. Accessibility: Plain language. Gaps are marked in words, not by colour alone. Accountability: A named officer approves. The AI’s draft is kept. Approval is refused until every gap is filled, every concern is checked and the officer adds their name. RESPONSIBLE AIThe officer stays in charge FairnessBoth sides go on therecord. The officerweighs them. BiasOnly facts from thenotes. No opinions,no guesses. AccessibilityPlain language. Gapsare marked in words,not by colour alone. AccountabilityA named officerapproves. The AI’sdraft is kept.Approval is refused until every gap is filled, every concern is checkedand the officer adds their name. Lightning RoD: AI Hackathon presentation · public summaryThomas Eapen · 2 / 3

Responsible AI

The officer stays in charge

Fairness

Both sides go on the record. The officer weighs them.

Bias

Only facts from the notes. No opinions, no guesses.

Accessibility

Plain language. Gaps are marked in words, not by colour alone.

Accountability

A named officer approves. The AI’s draft is kept.

Approval is refused until every gap is filled, every concern is checked and the officer adds their name.

Built in plain English, tested to breakIn plain English: The tool was described in plain English. An AI coding assistant wrote the code, and it was tested. Public rules: Every choice the AI makes follows public rules: the Employment Insurance Act, its Regulations and the published guidance on benefit decisions. The decision stays with the officer. Tested to break: 1. Fictional cases, most with traps; 2. The AI drafts, then reviews it; 3. The program checks the draft; a second AI call grades it; 4. Each miss becomes a new rule or check The officer decides. Lightning RoD writes it down. HOW WE BUILT ITBuilt in plain English, tested to break THE LANGUAGEIn plain EnglishThe tool was described in plainEnglish. An AI coding assistantwrote the code, and it was tested. THE CHOICESPublic rulesEvery choice the AI makes followspublic rules: the EmploymentInsurance Act, its Regulations andthe published guidance on benefitdecisions. The decision stayswith the officer. THE TESTSTested to break1Fictional cases, most with traps2The AI drafts, then reviews it3The program checks the draft;a second AI call grades it4Each miss becomes a new ruleor checkThe officer decides. Lightning RoD writes it down. Lightning RoD: AI Hackathon presentation · public summaryThomas Eapen · 3 / 3

How we built it

Built in plain English, tested to break

The language

In plain English

The tool was described in plain English. An AI coding assistant wrote the code, and it was tested.

The choices

Public rules

Every choice the AI makes follows public rules: the Employment Insurance Act, its Regulations and the published guidance on benefit decisions. The decision stays with the officer.

The tests

Tested to break

  1. Fictional cases, most with traps
  2. The AI drafts, then reviews it
  3. The program checks the draft; a second AI call grades it
  4. Each miss becomes a new rule or check

The officer decides. Lightning RoD writes it down.

Written for this site from parts of the presentation. The slides as presented are not published, and no case, real or fictional, is shown here.