← Araib Karim

AI work

Two AI systems I designed and led at Amazon, small enough to try.

Getting an AI to produce an answer is the easy part. Making it safe enough to act on real orders and real customer tickets is the work. These two demos show how I did that. Both use made-up data.

  1. 01Evidence grounding: an order platform that checks its own work ↓
  2. 02On-call agent: tickets resolved while the team sleeps ↓

Project 1 of 2 · Order platform

Evidence grounding

Customers send purchase orders as PDFs, and an AI turns them into real orders. A wrong value costs money, so nothing the model says is trusted until it is checked against the document.

Try it

Pick a field on the right to see where it came from in the document. Then try Deliver by, the one the AI made up.

PURCHASE ORDER
PO Number: PO-48213
Date: 14 Oct 2026
Buyer: Northwind Supply Co.
Ship to: 2200 Harbor Rd, Austin, TX 78701
40 × Safety gloves, size L @ $6.50
12 × LED work light, 20W @ $31.00
Total: $632.00
Notes: Deliver to dock B.

Extracted by the model

Illustrative demo with made-up data. The real platform took customer onboarding from 10 weeks to under 1 and grew from 2 customers to more than 25.

The three checks behind it

  • Rule checksDoes the answer make sense?For example, quantities times prices must add up to the total.
  • Evidence groundingIs it really in the document?Every value must be traced back to the source PDF. That is what the demo above shows.
  • Human reviewWhen unsure, ask a personLow-confidence or unfounded values go to a review queue, never straight into an order.

Project 2 of 2 · Operations

On-call agent

A separate system from the one above. This agent works the support ticket queue 24/7. It remembers past tickets, and it uses a tool I built on MCP to pull order details, logs and system data.

Try it

Pick a ticket to watch the agent work through it, step by step.

    Illustrative tickets with made-up data. The real agent cut the ticket queue from 70+ to 10–15, about 80%.