Skip to content

Allow this site to count your visits with Yandex Metrica? Metrica will store cookies in your browser to do this. How the site uses cookies

VC
RU
Case studies · 22 case studies and 11 ready-made solutions

My projects, with numbers before and after

Each case study covers the task, the solution, the technologies and the numbers before and after. Case studies and ready-made solutions fall into four groups by service. Some projects are under NDA, and I can talk about them on a call.

A · 2 case studies and 3 ready-made solutions

Talking to customers

Voice agents, round-the-clock booking, reviews, replies to missed calls and an AI sales assistant on the website. For business owners whose enquiries go unanswered after 7 PM.

Please note

I haven't yet built a voice agent, online booking or review collection for a client in these niches: salons, dental clinics and car repair shops. These are three ready-made solutions: I've worked out each one step by step, and GPT-4o, GigaChat and ElevenLabs voices already run in my other projects. The figures on these three cards are estimates for a typical business. The first 3–5 clients get 30% off and an extended 90-day warranty. In return, I ask for permission to describe the project without naming the company.

Round-the-clock booking Ready-made solution
Beauty salons and salon chains

A voice agent that answers salon calls day and night and books clients into YClients or another booking system

It answers calls at any hour, books clients into YClients or another booking system with an API, and sends an SMS confirmation. It knows 14 typical scenarios, and if a question is tricky, it passes the call to a receptionist. The call runs on Vapi or Bland AI with an ElevenLabs voice, and GPT-4o or Claude leads the conversation. Client data is stored where your country's law requires.

estimate: calls answered
64% 97%
Vapi / Bland AIElevenLabsGPT-4o / ClaudeYClients APIFastAPI
Read the case study
Related article AI receptionist and booking chatbot: where it breaks
Booking by urgency Ready-made solution
Private dental clinics

A voice agent for dental clinics that spots acute pain and books patients into the schedule

A patient calls or writes on WhatsApp, and the agent asks eight questions to work out how soon they need a dentist: straight away, within 24 hours or at a routine appointment. Then it books the patient into IDENT, Dentrix, Open Dental or another system with an API. The call runs on Vapi or Bland AI with an ElevenLabs voice. Patient data is stored where local law requires, and the AI model, GPT-4o or Claude, receives only the age group, sex and symptoms.

estimate: time to a booking for acute pain
47 min 4 min
Vapi / Bland AIElevenLabsGPT-4o / Clauden8nBooking system API
Read the case study
Related article Personal data and LLMs: 152-FZ and ChatGPT
Reviews and reputation Ready-made solution
Auto repair shop with 5 lifts · estimate

Reviews on Yandex Maps and 2GIS: the shop asks every customer for a review, and AI helps draft the replies

Eighteen hours after a work order is closed in Alpha-Auto (1C), every customer gets the same WhatsApp message: a thank-you, a question about how it went, and links to Yandex Maps and 2GIS. Customers get nothing for a review. The owner hears about a low rating right away and calls the customer to put things right. For each new review GigaChat drafts 2–3 replies, and the manager picks one and publishes it.

estimate: map rating in 4 months, if new reviews average about 4.8
4.1 4.7
n8nWAPI.cloudAlpha-Auto (1C)GigaChatMetabase
Read the case study
Support · telephony
E-commerce, US customers, support in another time zone

After a missed call, the customer gets an SMS with the real ship date

The phone system reports a missed call, and the service finds the customer and their orders in two Shopify stores and sends an SMS with the real ship date. In the CRM the call lands on a “Missed calls” page with the customer's orders and a reply ready to send.

~180 missed calls a month
no reply auto-reply
PythonWebhooksHMACSMS APIShopify APILLM
Read the case study
Related article AI receptionist and booking chatbot: where it breaks
AI sales assistant
Retail, brand website for outdoor gear

A store's own AI sales assistant won a blind test 34 : 12

Built for the brand's Russian site on Russian language models, so visitor data stays in Russia. It answers from a base of 562 product facts (RAG): the text appears as it's written, and product cards follow. Where a mistake would cost the customer money, the answer is set in code. Each lead reaches the sales manager with the whole conversation.

questions won out of 49 in a blind test: paid widget → own assistant
12 34
FastAPIRAGYandexGPTGigaChatSpeechKitTurnstilesystemd
Read the case study
Related article AI receptionist and booking chatbot: where it breaks
B · 8 case studies and 3 ready-made solutions

Process automation

Handling enquiries, marketing automation, finding new clients and tracking which ads bring in orders. For heads of sales and marketing.

Please note

The landing page that checks leads by tax ID, the cold email setup and server-side conversion tracking are ready-made solutions. I've worked them out step by step but haven't built them for a client yet, so the figures on these cards are estimates for a typical company. Sending sales to Keitaro and purchases to Facebook straight from the server already runs in my project for two Shopify stores.

Marketing automation
Ad agency (Facebook and Yandex)

Weekly client reports that put themselves together

Every Monday a workflow in n8n, an automation tool, pulls the numbers for all eight clients from Facebook, Yandex Direct, Yandex Metrica and Google Analytics. AI comments on sharp changes, and by 09:00 each client gets a Notion page or a PDF.

manual report building
4–6 h/wk 0 h/wk
n8nFB APIYandex.DirectClaudeNotion
Read the case study
Related article What AI implementation costs: rates and six estimates
AI · video ads · lip-sync
Performance marketing, ads in several languages

From speech to a finished ad: transcription, translation, a new voice-over and lip-sync

Whisper transcribes the speech, GPT-4o translates the script and rewrites it for the market, ElevenLabs recreates the speaker's voice with the right accent, and wav2lip matches the lip movement. One actress's video reaches 5+ countries with no reshoots.

per version of an ad, without lip-sync
4–6 h ~30 min
WhisperGPT-4oElevenLabswav2lipffmpegPython
Read the case study
Data collection · 30+ markets
Ad research

GMBSPYLAB: from the Facebook Ad Library to a Telegram feed

Async Python and Postgres 16. Every day 30–75k ads come in from the public Facebook Ad Library, and each country gets its own topic in Telegram. The response format is checked at every step: if the source changes something, the session restarts and collection carries on. The server and storage cost $30–60 a month.

ads a day, before and after removing duplicates
30–75k 5–12k
PythonPlaywrightaiogram 3Postgres 16arqMinIO/R2
Read the case study
End-to-end analytics
E-commerce, US brand, two Shopify stores

End-to-end attribution for two Shopify stores: from click to order

My own service on the Python standard library receives Shopify order notifications and checks their signature (HMAC). It sends the sale to the tracker and the purchase to the Facebook Conversions API with the same event ID as the pixel, so conversions aren't counted twice. A separate database links click, customer and order.

extra revenue in the report
+25.7% 0
Python (stdlib)Shopify WebhooksKeitaroCAPIGoogle Ads APISQLite
Read the case study
Operations automation
E-commerce, US brand, goods shipped from China

Shopify orders land in the warehouse sheets by themselves

Order numbers, tracking numbers, statuses and cancellations are filled in every 30 minutes. An Apps Script receiver inserts rows mid-sheet and carries the formulas down, leaving manual edits intact. Checks and backups run on a schedule.

orders into the warehouse sheets
manual every 30 min
PythonApps ScriptSheets APIsystemdTelegram APIpytest
Read the case study
Related article Inventory sync between 1C, Wildberries and Ozon
Operations · file exchange
E-commerce, warehouse and shipping partner, two stores

File exchange with a shipping partner: costs and tracking numbers come back into the sheet

The daily file builds itself and looks just like the working sheet, row colours included. The partner's reply is matched by order number plus row suffix. Shipping cost lands in its own row without overwriting manual edits, and bundle tracking numbers are spread across rows. A separate check catches the risk of shipping an order twice.

wait for the daily file
~50 min ~15 min
PythonopenpyxlApps ScriptTelegram APIsystemdpytest
Read the case study
Revenue reconciliation · analytics
E-commerce, US brand, two Shopify stores

Reconciling revenue across three sources: an 8.4% gap explained down to 0.33%

The owner's working sheet, the warehouse book and the CRM each showed a different revenue figure. A one-off check explained the 8.4% gap in site revenue down to 0.33% and caught a broken year-over-year comparison. The check now runs daily with a 3% threshold, and if less than 70% of the money can be compared, it gives no verdict.

unexplained gap in site revenue
8.4% 0.33%
PythonShopify Admin APIGoogle Sheets APIsystemdpytestJSON snapshot
Read the case study
Open-source intelligence
Competitive intelligence · my own tool

Researching a competitor from open sources: what you can see from outside

12 steps using free sources: WHOIS, DNS, certificates (crt.sh), URLscan and the web archive. A script collects the data, and conclusions follow a separate set of rules. The main one: shared tools do not prove a shared team.

spent on data; one site takes 30–60 min
paid subscriptions $0
bashwhois / dig / opensslcrt.shURLscanWayback
Read the case study
Customer acquisition · landing page Ready-made solution
B2B companies that sell through a demo or a call

An Astro landing page that scores B2B leads by company tax ID

A multi-step form asks for the company's tax ID. Within seconds an n8n workflow pulls data about the company from Dadata and SPARK and puts the lead in group A, B, C or D. A and B reach sales in amoCRM with a company profile, C gets an email sequence, and D gets a polite no.

estimate: time sales spends screening leads each week
13 hrs 4 hrs
Astro 5Tailwind v4n8nDadataSPARKamoCRM
Read the case study
Finding clients by cold email Ready-made solution
B2B companies selling to mid-sized and large businesses

Finding B2B clients with cold emails written by AI

Five steps. Apollo picks senior decision-makers who fit the customer profile, Clay collects 4–7 facts about each company (revenue from SPARK, news, website technology), GPT-4o writes an email from those facts, Instantly sends it from eight mailboxes, and replies land in amoCRM. Sales reps keep the quality checks and the conversations.

estimate: 300 personal emails a week
50 hrs/wk spot checks
ApolloClayInstantlyGPT-4oClaudeamoCRM
Read the case study
Conversion tracking Ready-made solution
Ad agencies and in-house media buying teams

Conversions go to the ad platforms straight from the server, and the team's report builds itself

The Keitaro tracker collects clicks and conversions on your own server. An nginx layer with Lua removes duplicates, and events go to Facebook, TikTok and Yandex Metrica straight from the server. Metabase shows return on ad spend, and Telegram alerts flag losing ads.

estimate: morning reconciliation per media buyer
2–2.5 hrs 0 hrs
KeitaroFB CAPITikTok Eventsnginx + LuaMetabase
Read the case study
C · 7 case studies and 3 ready-made solutions

AI infrastructure and integrations

Connecting AI to 1C and other business systems, document processing, search over a company's knowledge base, models that run on your own computer, networking, AI security audits and process audits. For CTOs and heads of IT.

Please note

The shop process audit, the chatbot security audit and the foremen's bot are ready-made solutions. I've worked them out step by step but haven't delivered them for a client yet, so the figures on these cards are estimates. A Telegram bot with three access levels and files for import into 1C already run in my other projects.

B2B · Finance
Accounting at a government-contract company

Converter for Federal Treasury statements into 1C

Reads XML statements from the Federal Treasury's GIS system in both format versions, V3 and V4, and builds a 1CClientBankExchange file in Windows-1251 encoding for import into 1C.

to process one statement
2–3 h 5 sec
Pythonxml.etreeWin-12511C
Read the case study
Related article Telegram bot for 1C: access without the 1C client
Calibration · AI evals
Prediction markets · my own project

Forecasting events with several language models: calibration and Brier scoring

Five AI agents search open sources, a sixth reconciles their answers, and a calibration step adjusts the probability. Every forecast is saved before the outcome and Brier-scored after it. Across 13 checks on 203 resolved markets the quality bar was never passed, so trading with real money stayed switched off.

Brier score, first check → 203 markets; market 0.075
0.119 0.104
PythonDuckDBClaude Sonnet ×5 + OpusWeb searchPlatt scalingBrier scoresystemd
Read the case study
Local AI · Ollama · RAG
Trading, fintech (NDA) · MacBook Air M3, 24 GB

Language models analyze trading signals right on a laptop: Ollama and RAG on an M3

qwen2.5:14b is the main model, llama3.1:8b the backup, and LLaVA:7b reads chart screenshots; ChromaDB finds similar cases in the trade history. There are no fees to outside services, and the strategy stays on the owner's computer. Dry run on 100 signals: 76% of signals skipped, 0 rule violations.

query costs; a cloud model at 1,000 signals a day would cost up to $100
~$100/day $0
Ollamaqwen2.5:14bLLaVA:7bChromaDBPythonM3, 24 GB
Read the case study
AI assistant · sales figures
E-commerce, US brand, two Shopify stores

A sales-figures bot: six metrics, with the arithmetic checked by AI agents

A Telegram bot reports sales by country, year, product and US state through one shared metrics core. All the numbers are computed in code. An hour after release four AI agents reviewed the calculations and found four errors, and five guard tests now keep them from coming back.

product-revenue gap against the platform report
+8.54% 0
PythonShopify Admin APITelegram Bot APIGraphQLpytestJSONL
Read the case study
Knowledge base · RAG
E-commerce, distributed team: owner, contractors, ops

An AI assistant for company knowledge in Telegram

The bot answers from work chats, email, voice notes and documents, and quotes its sources in every answer. Access is split into three levels. Search moved out of memory into SQLite FTS5, so it needs 20 times less memory, and the bot runs on a server even when the laptop is off.

knowledge-base search (memory 1044 → 50 MB)
253 ms 13 ms
PythonSQLite FTS5BM25WhisperTelegram APIMCP
Read the case study
Related article Personal data and LLMs: 152-FZ and ChatGPT
Reliability · real-time systems
Fintech · my own project

Autonomous bots and backtesting: how to make them reliable

Ten services make decisions from a stream of external events. Data arrives over WebSocket the moment an event happens, and the fast path from event to action takes 250–350 ms. Each event type has its own watchdog, ideas are tested on real history, and a risk check stops the system without waiting for a person.

until a silent failure shows up
10 h 6 min
PythonasyncioWebSocketDuckDB / SQLitesystemdbacktesting
Read the case study
Networking · routing
My own infrastructure · networking

Selective traffic routing through my own server

Some external services go through a server I rent, and the rest of the internet goes direct. Routes are set by address ranges (AS), with per-domain DNS bindings where needed. The router is configured through its HTTP API, a snapshot of the settings is taken before every change, and changes can be rolled back.

goes through the tunnel
all traffic only what needs it
WireGuardPython (stdlib)Router HTTP APIDNSsystemd
Read the case study
Process audit Ready-made solution
Online shops selling on their own site, WB and Ozon

An online shop audit and the three fixes that, by my estimate, pay off first

In a two-week audit I work out where key staff spend their time and draw up a list of tasks with a payback estimate for each. For a typical shop, three fixes top the list: stock sync across 1C, Wildberries and Ozon, AI for first-line support, and closing documents (UPD) generated in 1C.

estimate: stock mismatch across 1C, WB and Ozon
7.4% 0.3%
n8n1C REST+ODataRetailCRMWB+Ozon APIGigaChat
Read the case study
Related article Inventory sync between 1C, Wildberries and Ozon
Security audit · 152-FZ Ready-made solution
Services with an AI chatbot · estimate for an online school

I stress-test AI chatbots and find the vulnerabilities before anyone exploits them

I attack a copy of the system following the OWASP LLM Top 10, a public list of typical threats to AI applications. I check where personal data goes and whether 152-FZ, Russia's personal data law, is followed. Every finding comes with a fix plan, and after the fixes I run the attacks again.

target at the re-test
critical vulnerabilities 0
OWASP LLMPromptfooGarakRebuffCloudflare WAF
Read the case study
Related article Personal data and LLMs: 152-FZ and ChatGPT
1C integration Ready-made solution
Construction and fit-out companies

A Telegram bot for foremen: stock, budgets and payments from 1C right on site

A foreman opens the Telegram bot, sees data from 1C UNF in two taps and confirms deliveries with a photo of the delivery note. The aiogram 3 bot reaches 1C's built-in HTTP services through a Node.js middle layer, over an encrypted WireGuard tunnel.

estimate: a foreman's calls to the office
6–8 / day 0.4 / day
1C UNFaiogram 3Node.jsWireGuardMinIO
Read the case study
Related article Telegram bot for 1C: access without the 1C client
D · 5 case studies and 2 ready-made solutions

Turnkey development

Websites with telephony and a map listing, mobile apps in Swift and SwiftUI, web apps that install on the phone like regular apps (PWA), playable HTML5 ads and complete web services (Next.js and FastAPI). Behind this: more than 20 iOS apps in Swift, both training projects and my own, engineering work since 2017, and my own platform with role-based access.

Please note

The online school app and the fitness studio web app are ready-made solutions. I've worked them out step by step but haven't built apps like these for a client yet, so the timelines on these cards are estimates.

Website · map listing · phone system
Switchboard and LV switchgear manufacturer, Kaliningrad

Website, map listing and cloud phone system for a switchboard manufacturer

The manufacturer worked on referrals and had no presence online. Now it has an Astro website with a GigaChat chat assistant and a Yandex Business listing. Calls go to a Rostelecom cloud phone system with a voice menu for three departments, and the office number has been moved into it. Backups are verified, and a summary arrives every evening.

incoming call before and after the PBX was set up
dropped first ring at 19.6 s
Astro 5nginxPython / FastAPIGigaChatRostelecom cloud PBXYandex BusinessIndexNow
Read the case study
3 interfaces · 228 API endpoints · 221 tests
My own ad projects · internal tool

A platform for launching ad campaigns: from a desktop app for myself to a web service

It started as a Tkinter desktop app for my own use and grew into a Telegram bot and a web app on Next.js, FastAPI, PostgreSQL and Celery, with user roles and an encrypted token store. In summer 2026 the web app caught up with the desktop version. Ad spend from two networks reaches the tracker in each network's time zone, and reports trace a click to the ad and follow sign-ups through to first payments.

to launch one campaign; 2–10 in one batch
2 h ~15 min
Next.js 14FastAPIPostgreSQLCeleryTkinterpython-telegram-botDocker
Read the case study
Playable ad · HTML5
Casual mobile game (under NDA)

A playable ad for a mobile game: install rate up 73%, each install 38% cheaper

A 5–10 second HTML5 mini-level with the game's real mechanics, built with Cocos2d-js. One codebase produces up to 18 variants for five ad networks: Facebook, TikTok, IronSource, AppLovin and Unity Ads.

cost per install
$1.85 $1.14
HTML5Cocos2d-jsPlayable SDKWebpackTypeScript
Read the case study
Internal tools
E-commerce, a small operations team

A custom CRM for an ops team: email, orders, calls and chargebacks on one screen

The shared inbox, live data from two Shopify stores, shipments, calls and returns sit in one customer card. Email is handled by my own module on top of IMAP, and AI drafts the replies. The card opens three times faster.

time to open a customer card
10.3 s ≈3 s
Python (stdlib)IMAP/SMTPShopify APISQLitenginxpytest
Read the case study
Related article Inventory sync between 1C, Wildberries and Ozon
Desktop app
My own tool

A desktop app builds a promo page from a brief or a link

A Python and tkinter app builds the page from a Bootstrap or Astro template and a brief, or takes the content of an existing page by its URL. It finds the logo, favicon and background in the markup and uploads the finished archive to the host by itself.

to build a page: one run replaces manual markup edits
10 steps 1 run
PythontkinterPillowHTML/CSSREST API
Read the case study
iOS · first version of the app Ready-made solution
Online schools with live lessons

An app for an online school that reminds students about lessons and helps them practise pronunciation

A SwiftUI app for iOS 17 and later reminds students about each class a day, an hour and 10 minutes ahead, checks pronunciation right on the phone and shows progress. The trainer subscription is paid through the App Store, live lessons through YooKassa, and the backend runs on Firebase.

estimate: time to release in the App Store
kick-off ~9 weeks
SwiftSwiftUISpeech FrameworkStoreKit 2YooKassaFirebase
Read the case study
Related article What AI implementation costs: rates and six estimates
PWA · home-screen install Ready-made solution
Fitness studios and studio chains

A fitness studio app that installs from the browser and updates without the App Store

A progressive web app (PWA) installs on the iPhone home screen straight from Safari, works offline and sends class reminders on iOS 16.4 and later. Booking runs through YClients and memberships are paid through YooKassa. Updates ship without App Store review.

estimate: time to the first working version
kick-off 4–5 weeks
AstroWorkbox 7IndexedDBWebPushYClientsYooKassa
Read the case study
Related article What AI implementation costs: rates and six estimates
Behind the scenes

I'll show you the other projects on a call

Not every project can be published. On a free 15-minute call I'll show you how 2–3 unpublished projects close to your task are built, with numbers and technical details. Client names stay private.

Book a free call
Ready to start?

Start with the 5,000 ₽ audit: a written report and a payback estimate

I'll tell you what to automate first, when it will pay for itself and whether your task needs AI at all. Sometimes it doesn't.

Book an audit Message me on Telegram

Or just ask a question. During business hours I reply within 2 hours