It swallows the data
Sales, stock, pricing, ad spend, competitor listings and what shoppers search for — pulled from every marketplace into one place.
Home / How it works
Swallow the data. Spot the leaks. Fix them — on a leash.
Kafal is building Endcap as a pipeline, not a chatbot on a dashboard: connect, normalize, measure, then an AI layer that detects, diagnoses, recommends and, inside limits you set, acts.
It spots the leaks
“This SKU will stock out in Mirdif in three days.” “Visibility on this search term dropped 40%.” Only what actually matters.
It acts
Rather than leaving you to fix the spreadsheet, agents take the corrective action themselves — adjusting bids, flagging replenishment, fixing listings — inside your caps.
From the shelf to the client
The agent in the middle.
Same six platforms. Choose one to see what Endcap will actually read — then watch the insight leave the other side.
Six platforms in
Amazon.ae and Amazon.sa — SP-API and the Ads API. Orders, inventory, listings, search queries, bids and budgets. The richest source in the category, and the first place a write-back agent will run.
The agent in the middle
Endcap AI agent
Classical models find the anomaly and size it. The language model explains it in a paragraph a category manager can act on in five seconds — and never invents a number.
DetectDiagnoseRecommend with AED impactAct inside your caps
1
WhatsApp and Slack
The alert lands where the team already is, with cause and AED impact attached.
2
The morning queue
Category, KAM, supply and CFO each open a different first screen.
3
Suggest → Approve → Auto
Nothing writes back until the brand has seen the rule be right.
The AI layer
Four tiers, and two rules we will not break.
Models compute. The language model explains.
Never let an LLM generate a number a customer will act on. It queries the warehouse through tools and narrates what comes back. Every hallucinated figure in a weekly review costs more trust than ten good insights earn.
Autonomy is earned, not switched on.
Every action starts in Suggest, graduates to Approve after the brand has seen it be right, and only then goes to Auto — and even then only within caps. Per action type. Per brand. Never globally.
What changed?
Classical anomaly detection over every time series, with seasonality so Ramadan does not look like a crisis. Plus a rules engine for known failures: buy-box lost, listing suppressed, ad budget exhausted before noon.
Why?
A signal on its own creates work. Endcap walks the causal chain through the normalised model and turns it into a paragraph a human can act on in five seconds.
What is it worth?
Raise the bid on “face wash” from AED 18 to 24 → estimated +AED 21k GMV over 30 days, 82% confidence, based on 14 comparable moves. Without the number, a recommendation is an opinion.
Do it
A deterministic workflow engine. The model chooses which pre-approved action to run. The engine runs it. Logged, reversible, kill-switch ready.
Guardrails
Trust is the product. We build it before the first write.
Suggest
Always the starting mode
The agent proposes. A human decides. Nothing writes back yet.
Approve
One click, logged
After the brand has been doing that click for a while, the same action can move to Approve.
Auto
Only after two more things
A ninety-day backtest — we replay history and show what the rule would have done — and a live holdout of SKUs the rule is never allowed to touch.
Never auto
Price, and the ERP
Price changes and anything written to your ERP stay behind a human, permanently. There is a kill switch. There is one-click rollback.
Ask Endcap
A colleague, not a report.
A text box where a KAM types “why did our face-wash sales drop in Dubai Marina last week” and gets the answer, the chart and the suggested fix. It is a thin layer over diagnosis. We build it once the numbers underneath are solid — not before, because a conversational layer over shaky data just makes wrong answers sound confident.
Endcap · Aurex — Hand wash low stock in Mirdif. Suggested: top up 120 units. Reply 1 to approve, 2 to snooze.
Weekly review
The Monday deck writes itself from the week’s numbers, with commentary drafted from the same warehouse the console uses.
Audit trail
Every automated action logged and attributable. UAE PDPL, marketplace terms of service, and a record finance can open.
Design-partner build · Dubai
See a causal chain, live.
Open an alert, walk the diagnosis, approve a bid, and watch realised impact appear against the estimate.
