Astarmo · Decision infrastructure for e-commerce
AI at every decision.
A human before every change.
Six tools say what happened.
None says what to do.
A store generates data on sales, costs, prices, stock, products, customers, campaigns, returns and conversion. Existing tools display it and compute it well.
The missing layer is the work between “something happened” and “here is a specific decision, ready for approval”. Someone has to do that work: connect the right facts, recognize a situation that needs attention, check the context, apply the store’s rules, prepare a proposal, show the evidence, and carry the approved change through to execution and outcome.
Astarmo takes that work on. The operator keeps authority over every change — and stops assembling each decision from scratch.
Problem
Three leaks that only show up in the books.
If at all.
Leak 01
The ROAS that lies
The ad platform reports campaign revenue — it doesn’t know the cost of goods and can’t. Nobody subtracts it, so a campaign with a 600% ROAS can lose money while the report looks like a success.
Astarmo
Breakeven ROAS per campaign, computed from purchase costs and revenue measured by its own tag. A campaign below the threshold gets a proposal to cut — backed by a number, not a hunch.
Leak 02
The price that fell behind
The purchase cost went up; the price and the promotion stayed where they were. The product sells below breakeven for weeks, because nobody recalculates margin item by item.
Astarmo
Margin computed from real costs at every sync. The signal goes out the same day, not in a monthly report.
Leak 03
The decline noticed last
Conversion slides week after week. It only shows in the monthly result — once the money has already leaked and there is nothing left to save.
Astarmo
A baseline from the store’s own history. A deviation becomes a signal before the month closes.
These decisions get made every day anyway. Today: by hand, in the dark, with no memory.
How Astarmo detects these three situations and where the numbers come fromConversation
A simple question.
An analysis no spreadsheet can do.
A dashboard answers “what happened”. The operator needs an answer to “when should this be done” — and that takes reading orders, costs, signals, executed decisions and rules all at once. Astarmo asks about all of it on its own and answers with a date, not a platitude — even when that date falls later than the operator hoped.
Operator
when to mark down gaming chairs without eating into margin?
Astarmo
After August 22, and on 31 of the 48 models. That’s when the supplier promotion expires, the purchase cost returns to 682.40 PLN, and the rest of the category climbs back above the 22.0% threshold the operator saved to memory.7
Why not today, and not the whole category
- 17 of the 48 models sit below the 22.0% threshold today — a markdown would push them deeper instead of closing the leak.12
- Deepest of all is the Vantar RX: its purchase cost rose by 38.00 PLN on July 11 and was reflected in neither the price nor the promotion.3
- The other 31 models have headroom — median margin 29.4%, so a 10% markdown stays above the threshold.45
- Traffic in the category is stable, so the margin drop is explained not by demand but by purchase cost.6
Consulted
- 1,240
- orders
- 48
- models in category
- 18
- cost lines
- 7
- store-model queries
- 3
- memory rules
- 41,812
- audit entries
- 0
- personal data
The conversation changes nothing. For anything to change in the store, the operator has to approve a proposal.
The store model this answer stands onAI
The best available model.
Zero executive power.
Astarmo runs on frontier language models — Claude today, swappable by design. The model’s provider and version are recorded in the fingerprint of every decision. When a better model appears, the system hires it. The infrastructure doesn’t age along with any single provider.
The model doesn’t get a summary of the store — it gets the keys to it and decides for itself what to ask: the margin on this product, what the operator rejected six months ago, whether a similar change has worked before and with what effect. The full surface of questions, not a prepackaged bundle of context.
It doesn’t define its own permissions. It doesn’t set its own limits. It doesn’t execute the change. The model may be brilliant at thinking — Astarmo doesn’t rely on it following the rules of its own accord. The rules are enforced by the system, outside the model.
The bill doesn’t depend on tokens.
The model is Astarmo’s cost, not the store’s — which leaves a free hand to reach for the best model available rather than the cheapest. The store pays a fixed monthly fee for a working system, not for usage.
The invoice doesn’t grow with how many times the model read the store’s history, how many proposals the operator rejected, or how many times the detectors recalculated margin. What grows is what the system has managed to compute and settle — not what it drew from the model provider.
- Bounds
- Every parameter from the model passes deterministic bounds validation. Rejection, never a silent correction.
- Personal data
- Name, email, address, IP and ad identifiers are stripped before every prompt. The model doesn’t even get the store’s identifier.
- Limit
- A cap on model spend, counted separately for each store. Once it’s exceeded, the AI layer goes quiet and the deterministic loop keeps running.
- Off switch
- The operator switches off the AI layer for their store with a single decision, and needs nobody’s sign-off to do it.
The model proposes every one of these decisions.
It executes none of them.
Full cognitive power · Zero executive power
Platform — one loop
One decision passes through six stations.
-
01
Signal
Code computes the fact — deterministically: the same data gives the same result. Whether the fact matters is for the model to judge.
Where a signal comes fromdetector margin_leak window 28 days basis orders + purchase costs result 3 SKUs below threshold
-
02
Evidence
Every proposal carries numbers and their origin. Where from, what period, how many.
From signal to a finished proposalmargin 24.1% -> 19.8% −4.3 pp period 2026-07-11 .. 2026-08-07 source orders 142 · costs 18 check store data, not a benchmark
-
03
Approval
No change touches the store without the operator’s approval. That’s architecture, not a setting.
The decision inbox and the dry runproposal -> operator -> store ________ the only way in model has no such path -
04
Execution
The write passes through a closed registry of gates. Cutting it off takes seconds.
Ports, guard_write and the kill switchguard_write ENV ok kill-switch off RLS ok port price.prestashop -
05
Measurement
Every executed decision gets a scorecard. Effects as intervals, labeled honestly.
The scorecard and an honest reading of the outcomemargin +2.1 pp CI [+0.4, +3.8] revenue +1,240 PLN CI [−310, +2,790] reading correlational, not causal
-
06
Memory
Approved rules become the law of the system. Enforced without exception.
Memory and a rule’s vetorule minimum margin 22.0% status enforced effect proposal −11.0% REJECTED
Principles
Four sentences that constrain Astarmo more tightly than any terms of service. Each one has its own test in the code and fails the deployment the moment it stops being true.
The evidence chain, entry by entry- Art. 1
The language model proposes. It never executes.
- Art. 2
Every AI event has a fingerprint and a trace in the chain.
- Art. 3
The operator’s rules can veto any proposal.
- Art. 4
When the system doesn’t know a number, it says “I don’t know”.
Nihil novi sine communi consensu · 1505
Show us a decision you make by hand today.
The first conversation starts from a real operating decision. We establish which facts it needs, where they come from, which rules apply, and what the path from signal to action looks like today. If Astarmo can shorten that path and give it structure, that’s where the rollout begins. One message opens the conversation: a description of the decision is all the inquiry needs, and a reply comes back within one business day.