Skip to main content
TECHIRELANDNATIONAL
AI
CHALLENGE2026

Akella inMotion

Blueprint

Part 02 of fourThe build.

Eighteen days. Twelve working days. One proof that must survive a clean browser, a regional table, and the question you least want to hear.

Build the risky claim. Simulate the plumbing. Label the seam. Freeze on Day 13.

18
calendar days
12
working days
13+
regional hubs
~1k
2026 registrations

Current 2026 scale from TechIreland material reviewed 22 August. Dates and organiser instructions can change, so the Challenge platform remains the source of truth.

The operating rule

A proof is not less serious than a product. It is more specific.

01

Build

The claim.

The number or moment that decides whether the idea is real.

02

Simulate

The plumbing.

Slow or irrelevant dependencies that add risk but no proof.

03

Label

Every seam.

The truth table that turns a shortcut into an honest prototype boundary.

Choose your route

Same standard of proof. Different controls.

No-code is not the absence of architecture. It moves architecture into screens, records, states, permissions, prompts, and platform connections. Pick the route that owns the proof core and let this page filter the workbench. Or let the mapper pick it from what you tick.

Mixed team? Choose the route that owns the proof core. Use the other route at a documented handoff, with one shared contract and one owner on each side. Not sure which owns it? The mapper will tell you.

The route workspace

Choose the work in front of you.

Open the playbook, proof boundary, schedule, resources, agent kit, or working files without losing your route. Copy the URL when somebody else needs the same panel.

Technical view

Route A workbench

One repository. One public URL. One walking skeleton before breadth.

Use this route when somebody can inspect source, logs, requests, and deployments. The advantage is control. The danger is spending the fortnight on architecture nobody sees.

Route A rule

One repository. One public URL. One walking skeleton before breadth.

01

Architecture

Draw the walking skeleton.

Outcome

Input reaches the proof core and a result reaches the screen on a deployed URL.

Workflow

Keep one app and one data store. Draw the request, model, tool, storage, timeout, and fallback path before adding services.

Prompt starter

Propose the smallest architecture that proves this claim. Name the largest unknown, one rejected alternative, and the one-day spike.

04-ARCHITECTURE-DECISION.md
02

Frontend

Build the money screen first.

Outcome

A judge can see the changed process without an architecture explanation.

Workflow

Start from a strong template, use three screens at most, and design for a projector. Replace blank spinners with a visible trace.

Prompt starter

Build only the central screen in the existing visual system. Include loading, empty, success, error, and reset states.

05-MONEY-SCREEN.md
03

Backend + API

Agree the contract before both sides exist.

Outcome

Frontend and backend can work in parallel against the same request, states, errors, and example response.

Workflow

Write the contract, stub it at the boundary, keep keys server-side, and make seed plus reset one command.

Prompt starter

Write the smallest typed API contract for this flow. Include examples, timeouts, retry rules, failure shape, and idempotency.

06-API-CONTRACT.md
04

AI + evals

Turn the claim into twenty cases.

Outcome

The model or workflow is judged against a fixed set, not the last example the team liked.

Workflow

Start deterministic, then one model call, then retrieval, then tools, then an agent loop only if the case set requires it.

Prompt starter

Create a balanced twenty-case eval set with expected properties, edge cases, pass threshold, latency, and cost fields.

08-EVAL-PLAN.md
05

Agents + tools

Add autonomy one permission at a time.

Outcome

Every tool has a narrow input schema, a stop condition, a trace, and a visible human control.

Workflow

Use a workflow when the steps are known. Use an agent only when the system must choose the next step from evidence.

Prompt starter

Define the agent role, allowed tools, stop conditions, approval points, output schema, and trace. Remove every optional tool.

09-AGENT-CONTRACT.md
06

Deploy + recover

Ship broken on Day 1.

Outcome

The public URL, logs, seed, reset, feature flags, error state, and backup path have all been exercised.

Workflow

Deploy the thin slice immediately. Test the cold path daily. Record the happy path before freeze.

Prompt starter

Audit the deployed proof path. Return PASS, FAIL, or NOT VERIFIED with exact evidence and the shortest recovery action.

10-DEMO-RUNBOOK.md
07

While you sleep

Schedule evidence, not authority.

Outcome

Codex, Claude Code, Gemini CLI, or CI leaves a reviewable report or isolated patch by morning.

Workflow

Allow tests, evals, link checks, accessibility reports, and bounded worktree patches. Forbid merge, deploy, schema change, purchase, or external messages.

Prompt starter

Run the listed checks in least privilege. Stop on ambiguity. Do not push, merge, deploy, mutate data, or reveal secrets.

12-OVERNIGHT-TASK.md

1 of 11

Next in the series

03. The room.

The Challenge is judged on the presentation of an idea, not on the product. The five beats, the word budget, the deck, the demo runbook, the landing page, and the eleven questions you will be asked.