An AI sales analyst for marketplaces, built on Hermes

THE TASK

Plenty of reports, but no idea what to do today

A seller on two marketplaces has dozens of reports: orders, buyouts, ads, stock, commissions. The numbers are there, but problems are spotted late:

a bestseller ran out of stock — noticed only when orders dropped;
the price went up, orders sagged, and the link was spotted a week later;
the ad bid went up, spend grew, but orders stayed the same;
the marketplace raised its commission — and some products quietly started selling at a loss.

What's needed is an analyst who looks at every number every day, finds the anomalies and says what to do. And answers a question in a minute rather than after an export to Excel.

HOW IT WORKS

A question in plain language — an answer with a number and an action

01

Ask it like a colleague

«What needs attention?», «Why did Ozon sales drop?», «What should we ship to the warehouse?» — in Telegram or the web chat. At 9 am the agent sends a digest on its own.

02

The agent takes numbers from the data

Nine MCP tools: totals, products, unit economics, daily trends, stock, ads, anomalies. Profit, ad cost ratio and days of stock are computed in code — the model doesn't recalculate them.

03

It answers with a reason and an action

Not «optimise your ads» but «ad cost ratio went from 8.9 to 23.8% with the same orders — review the bid». A person decides: the agent changes nothing in the seller accounts.

CAPABILITIES

What the agent does

01

Morning digest

Yesterday and the week in numbers, up to five «important today» items — each with an action and a number — and the good news. Readable on a phone in a minute.

02

Finds anomalies on its own

Rules in code compare the last week with the four before it: orders falling or growing, the effect of a price change, a jump in ad cost, falling buyouts, loss-making sales, a new commission, stock.

03

Tells season from a problem

When the whole category grows, that's the season, not the listing. When orders drop, it looks for the cause in the data: price, ads, views, conversion or an empty warehouse.

04

Calculates unit economics

What's left from one unit after commission, logistics, storage, ads and tax, and at what price the product breaks even.

05

Plans restocking

How many days the stock will last, when it runs out, how much to ship given the delivery time, and how many orders are already lost.

06

Reviews the ads

Spend, CTR, cost per click and ad cost ratio per campaign: you can see where the bid went up but orders didn't.

TRUST

Why the agent's numbers can be trusted

Code does the maths, not the model. Profit, margin, ad cost ratio, buyout rate, break-even price and the comparison with the previous period live in the MCP server. The model takes ready values and explains them. Charts in the web chat are built from the data too: the model only says what to show.
Rules find the anomalies. Growth and decline only count when they exceed random noise. Every finding comes with the numbers behind it and an estimate of money per week.
Read-only. The agent has no write tools at all, and the database is opened read-only. Terminal, files, browser, the internet and memory are off: the model sees nine analytics tools and its skills.
Data, not commands. Hermes marks tool results as untrusted, and the agent treats product and campaign names as data: it won't follow instructions hidden in them.
Every call is logged. Which tool, with which arguments, whether it succeeded and how many milliseconds it took.

IN NUMBERS

Tested, not promised

Situations in the data 8 built in — the rules find all of them unprompted and raise no false alarms

Morning digest 30–60 seconds and 5–7 data requests; every item comes with an action and a number

Automated tests 65: calculations, the MCP protocol, the web chat and the live chain in containers on a mock model

Tools 9 MCP tools — not one of them writes

Deployment one command: data, Hermes, the MCP server, the web chat and the scheduled digest

AI model ChatGPT subscription, no API billing; the same Hermes version as the AI finance agent

FOUND DURING DEVELOPMENT

Where an analyst goes wrong if the maths is sloppy

Three traps caught by the tests. Each one produced a confident but wrong answer.

Noise looked like growth. For a product with four orders a day, «+49% in a week» is plain chance. Now a change only counts if it exceeds random variation.
A price rise hid inside the average. The price went up 10 days ago — some days after the rise fell into the baseline, and the drop in orders got diluted. Now the price effect is measured from the day of the change: before and after.
Examples in the instructions gave the answer away. Numbers in a skill's example matched the real ones — the model «knew» the answer without looking at the data. The examples were made neutral, and the tests check answers against a direct calculation.

How it's checked

A mock model calls the tools through the real Hermes and builds its answer from what they return. The test checks the answer's numbers against a direct calculation on the same database and verifies that the model sees only the analytics tools.

Want an analyst like this for your seller accounts?

I'll connect the agent to your Wildberries and Ozon accounts: data via the marketplace APIs, cost prices from your spreadsheet, the digest in Telegram. We can start with one marketplace. Message me on Telegram and we'll go through the task.

Message on Telegram dnovoselov@bk.ru

I reply within 3 hours on weekdays

I work under a contract and an NDA. My price ranges are in the Pricing section.