INTERVIEW WALKTHROUGH / FICTIONAL ORDERS
Walk through Field Relay in five minutes.
Find a difference. Inspect the original rows. Then explain what happens when an order is saved but its confirmation never arrives.
This is an AI-assisted portfolio prototype. The supply-chain coordinator below is hypothetical. The website works with files in your browser; the separate HTTP experiment was run locally.
Open the reconciliation desk01 / TRY THE WORKING DESK
What changed between these exports?
Imagine a coordinator comparing a supplier export with the team’s sheet before a handoff. Some values have changed. An order is missing. An ID appears twice.
- Open the desk. The fictional exports and their field mappings are already loaded. Choose Reload fictional example to start over, or Try a fictional order handoff after clearing the example.
- Select a finding. Read the baseline and candidate rows instead of guessing which version is right.
- Record a review decision with a reason. Duplicate and blank IDs remain ambiguous. Download the evidence before closing the tab.
“Local” means these files stay in the current browser tab. Reviewing a difference does not send an order or update another system.
02 / EXPLAIN UNCERTAIN DELIVERY
A timeout does not tell you whether the order was saved.
The coordinator submits an approved order. The supplier saves it, but the reply is lost. Sending a new order could create a duplicate. Assuming nothing happened could hide accepted work.
The local relay preserves decisions through restarts, limits retries and holds conflicting evidence for investigation. Approval before a send and revocation after a send are different events: an already admitted request can still commit.
What can I actually operate in this website?
The reconciliation desk and the Delivery simulator lab run in the browser. The simulator is a separate fictional exercise with outage, retry and replay controls. It consumes no reconciliation result. The Python/SQLite HTTP experiment shown below is recorded evidence, not a remotely connected backend.
03 / CHECK THE EVIDENCE
Recorded local HTTP experiment
5 scenarios passed; 2 exact supplier orders verified. These are measured results from fictional local fixtures, including injected failures. They are not customer outcomes.
| Scenario | Result | Verified orders for this scenario |
|---|---|---|
| Supplier saves the order, confirmation is lost | passed | 1 |
| Restart and repeat the verified order | passed | 1 |
| Change the amount under the same order ID | passed | 1 |
| Revoke approval while waiting to retry | passed | 0 |
| Revoke after the send has been admitted | passed | 1 |
Recorded 15 October 2026. Source 10733d1; clean source: true. Scenario counts refer to separate checks and must not be added together as unique orders.
Inspect the reports and failed attempts. An earlier run stopped before admission and did not retain its exact cause. The repair separates relay startup from the supplier wait and preserves safe failure codes. Historical failures remain visible.
04 / SAY IT IN YOUR OWN WORDS
The answer you should be able to defend.
“I built Field Relay with AI coding assistance to explore an operations problem: how do you understand changes between exports, and recover when an order request times out? The desk shows source rows. The separate local lab keeps the approved order fixed, uses one stable identity, and checks what the supplier actually saved. I inject failures and inspect the results. It is a learning prototype using fictional data.”
Be ready to explain why a successful response is only a claim, why a changed amount needs fresh approval, and why bounded retries can leave an unresolved exception. Describe your own decisions and learning honestly.
Where could Claude fit later?
A proposed extension could turn a messy supplier message into an order proposal. Application code would validate it, a person would approve exact fields, and the deterministic relay would handle delivery. A model should not infer permission or invent a delivery status. No Claude call exists in this prototype.