Agents and bots that run in a live store, not in a demo.

An AI assistant on the store catalog, a booking bot for a solo pro, an agent that checks product photos and rolls back its own fixes. Every reply passes code checks before a customer sees it, and each day's dialogs are scored so I can fix the scripts.

The first assistant answers shoppers of an appliance store every day: 328 messages a day.

Test dialogue with the AI consultant: picking a fridge under 60 cm from the tt.ua catalogue, warranty, delivery to the floor
Agents and bots

AI consultant for the Technotochka appliance store chain: answers shoppers in chat and does not invent prices

A bot in the store's Telegram and Viber: picks appliances from tt.ua, quotes delivery, checks orders, hands leads to a manager. Runs the chats alone at night.

8,059tests green on every production release

A product on a grey background: the agent rolled back the exposure fix, replaced the background and accepted the photo. Below the result, the trace of steps3:37
Agents and bots

ListingLens: an agent that checks product photos before they go live on a marketplace

An agent on OpenCV 5 checks a product photo, fixes what it can, re-measures the fix and rolls back what did not help. 89.6% correct decisions on 1,768 photos.

89.6%correct decisions on 1,768 hard photos versus 85.2% without verification

open brief

Afterglow home page: 'Your shop sees what sold. Afterglow sees what walked out', one real request from start to the restock list
Agents and bots

Afterglow: an AI agent for a record and book shop that sees what walked out the door

A shopper types 'gift for my brother, loves Joy Division, under $40'. The agent asks Qloo, picks three things off the shelf; the missing ones go to the owner.

25 / 9wins and losses against Claude without Qloo on 50 requests we never tuned on

open brief

Cover of the setup guide: 'The client writes at night. You sleep. The bot has already booked them for tomorrow'
Agents and bots

Booking bot for solo professionals: the client writes at night, the bot has already booked them for tomorrow

A pro connects their own Telegram bot in 5 minutes: clients pick a service, day and hour with buttons, the pro is notified, the client reminded. $10 a month.

2 mo 20 dthe service ran without a single restart at the time of the check

The Well-Fed Hamster card: market cap, BTC and USDT dominance, fear and greed index, altcoin season index, seven assets
Agents and bots

Well-Fed Hamster: a daily market snapshot for a crypto community, assembled by an AI agent

Each morning the agent pulls market cap, dominance, indexes and seven assets from TradingView and CryptoRank, renders a card, hands it over. 47 days in a row.

47cards in 47 days, not a single day missed

Echo Show screen: "Not for Saros 10R", the store's quote and a part the store actually lists for this robot2:56
Agents and bots

Shop Counter: Alexa+ remembers your devices and answers "will it fit?" only in the store's own words

MCP server for Alexa+: one device closet for every store, a yes only when the store lists your exact model, the quote on screen. 81 of 81 with nothing made up.

81 of 81shopper requests with no made-up fact, checked against the store's raw feed

open brief

Left: a bot draft with four false promises, right: the corrected reply sent to the customer, below: the list of violations3:20
Agents and bots

ShopGuard: an honesty guard for shop chatbots on open models

ShopGuard checks every promise a shop bot makes against the shop's data and the web before the customer sees it. Made-up promises: 8 of 49 became 0 of 49.

0 of 49made-up promises reached the customer, down from 8 of 49 with a pushy bot before the guard

open brief

The Launch board in Stackline: three lists, numbered cards, overdue in red, "waits on #4", prompts for Claude
Agents and bots

Stackline: a team task board run from Claude and ChatGPT

One sentence becomes a board: cards with dates and a chain of blockers. Run it from the page or from Claude and ChatGPT over MCP. 47 tests.

47tests in the second version, live check 11 of 11 on the real DataK3 database

open brief

What every bot has

Code checks on every reply. price without data, "in stock", invented actions, phone number, delivery city. On a violation the model rewrites, otherwise the customer gets a safe text and a human.

Daily review of live dialogs. instruments instead of guesses: escalation share, rewrite share, night requests, catalog coverage.

Release only on green tests. and on the owner's word, with a rollback tag on each release.

What it does not do

Discounts and haggling. the bot never promises a discount a manager has not approved.

Stock availability. the bot says neither "in stock" nor "out of stock" because it has no access to inventory. The owner's decision.

A perfect first version. in the first week the bot talks not quite the way you want. Someone has to look at the dialogs once a day.

What I need from the store

The catalog. a website, a CMS export (Shopify, WooCommerce, OpenCart) or at least an Excel file.

Written rules. delivery, warranty, installments, what may be promised and what may not.

One person. who looks at the dialogs once a day and says where the bot said something silly.