//01
One person — one timeline — one point of responsibility
You talk directly to the person writing the code. Decisions and fixes take hours, not days of manager-to-manager relay. No markup for an office or a "project team".
50+
AI bots and agents shipped to production — running for clients right now
4
apps live on the App Store, a fifth in review — check them in one click
3–5K
documents per month processed by a system running for a current client
250+
production tasks shipped in a year; case-study ROI up to 420%, payback in as little as 3 months
// AN_HONEST_ANSWER_TO_THE_MAIN_QUESTION
Why one engineer and not an agency
//02
The AI never makes up prices
Prices and business rules come only from your database and live in plain, verifiable code. Money and access are never entrusted to a neural network — which is why my systems don't fall apart on real customers.
//03
Responsible after launch
Monitoring, backups, a rollback plan for every change, and an honest report of "what was verified / what wasn't". My systems run for years and grow new features without being rewritten from scratch.
Got a task in mind? Tell me about it — I'll say honestly whether you need AI at all.
// SERVICES
AI implementation: what I do
Six service lines. Each one lists the real pain points clients come to me with — and how I solve them. All examples are from real projects, anonymized.
AI sales agent in messengers and on your website
// SOUND_FAMILIAR?
- Inquiries arrive at night and on weekends — by morning the customer has already bought from a competitor.
- Your managers take an hour to reply, and they explain instead of selling.
- You tried a "button bot" — customers get annoyed and abandon the chat.
- You're afraid to trust AI with prices and terms: if it lies, you're the one cleaning up.
// HOW_I_SOLVE_IT
- A natural conversational consultant with a managed sales funnel: the stages "greeting → discovery → recommendation → closing" are controlled by code, not by the model's mood.
- A catalog of 13,000+ SKUs with zero hallucinations: prices, specs and photos come only from your product feed — the model simply has no way to "misremember" them.
- Legally sensitive wording (prices, discounts, installments, mortgages) is locked down with forbidden-phrase lists.
- Before launch — a stress test with virtual shoppers: 15 AI "customers" run through 16 scenarios; a typical iteration catches ~11 bugs before any real person sees the bot.
- Hot leads are handed to a manager with the full conversation context.
// FROM_REAL_PROJECTS: a 13,000+ SKU catalog · rebuilding a "button bot" into a real consultant · a real-estate developer's bot with legally vetted wording on prices and mortgages
// FORMAT: pilot → project (from ₽300,000) → ongoing support
AI inside your CRM: chats, calls and quality control
// SOUND_FAMILIAR?
- Managers drown in repetitive questions and never get to the hot leads.
- You don't know what's happening in chats and calls until the customer is already gone.
- "Quality control" means spot-listening to 5 calls a month.
- Call recordings can't go to third-party clouds — your security team says no.
// HOW_I_SOLVE_IT
- An AI autopilot in amoCRM / Bitrix24 chats: it answers only from your company's knowledge base; if the answer isn't there, it brings in a human instead of making things up.
- Speech analytics on 100% of calls with on-premise speech recognition: audio never leaves your server; every call gets a transcript and a score against a checklist built from your own playbooks, results land in the CRM automatically.
- A voice AI agent that makes calls and holds a real-time conversation — outbound campaigns and win-back of lost customers.
- Live dashboards on top of your CRM: the funnel and manager leaderboards in real time, up to a TV screen in the office. Production CRMs are touched only with backups in place.
// FROM_REAL_PROJECTS: chat autopilot in amoCRM · speech analytics for a service group's operations · data → decisions
// FORMAT: pilot in one department → scale-up (project from ₽300,000) → ongoing support
Documents that process themselves: invoices, receipts, contracts
// SOUND_FAMILIAR?
- Accounting spends 5–10 minutes retyping every delivery note into the ERP.
- Errors in details and totals surface at month-end close — and cost real money.
- The document volume keeps growing, and hiring people just to retype paper feels wrong.
- Documents get lost in email and messengers; the history of contract decisions lives in your lawyer's head.
// HOW_I_SOLVE_IT
- A pipeline of "scan / photo / PDF → structured data → your ERP (1C)", running in production at 3,000–5,000 documents a month: 10–30 seconds per document, 99%+ accuracy, 95% fewer manual-entry errors.
- Protection against AI hallucinations: an anchor check that verifies it's the right document, plus rejection of broken or rotated scans.
- The model is chosen by benchmarking on your documents — my own test rig, not vendor marketing; models are swapped with zero downtime.
- A contract pipeline: drafts and reviews based on your legal team's historical practice.
- For restricted environments — a fully local option with no external AI services.
// FROM_REAL_PROJECTS: ROI 385% for a distributor · ROI 310% in construction · re-processing complex documents for an international broker
// FORMAT: pilot on your own stack of documents → project (from ₽300,000) → ongoing support
Portal, dashboards and forecasts: a single source of truth
// SOUND_FAMILIAR?
- Records live in dozens of Excel files — every department has its own version of the truth.
- The P&L appears once a month, cobbled together by hand — decisions are made on stale numbers.
- Your ERP is locked inside the security perimeter: no safe way to get data out.
- Meeting decisions dissolve in chats and memories; when an employee leaves, the history leaves with them.
// HOW_I_SOLVE_IT
- A web portal built around your processes: entity cards, registries, multi-step approvals, role-based access; several legal entities of a group in one portal with no data crossover.
- A "narrow bridge" to your ERP: data leaves the closed perimeter and returns without exposing the ERP itself; an auto-parser understands 7 different "human" spreadsheet formats.
- A live P&L: your management Excel moves to the web without losing a single figure, with honest "actuals vs preliminary" labeling.
- Forecasts on your data: 94% demand-forecast accuracy for a 15-store retail chain — purchasing without stockouts or overstock.
- Company memory: every meeting automatically becomes an executive brief delivered to the right people in Telegram; customer history is recoverable even from archives — proven on 53,000 emails.
// FROM_REAL_PROJECTS: forecasting for 15 supermarkets · a portal for a nationwide service group · a live owner's P&L · meeting digests for three executives
// FORMAT: pilot on one process → portal (from ₽300,000) → growth by subscription
B2B lead generation and customer win-back
// SOUND_FAMILIAR?
- Purchased contact lists are dead; emails to them go straight to spam.
- Building a dossier on one company before reaching out takes half a day.
- Cold emails are generic — nobody replies.
- Your own base of churned customers sits dead in the ERP.
// HOW_I_SOLVE_IT
- A lead radar over open government registries: 400,000+ companies filtered by your criteria, auto-enriched with contacts and refreshed automatically.
- An outreach pipeline: an online dossier on each company → a personalized draft written by AI → approval by a human → send. A human is always in the loop.
- A news radar: business events scored for "is this a reason to reach out".
- Win-back of your own base: churned customers exported from the ERP (proven on 10,000+ records), segmented, and re-engaged — up to a voice AI agent that makes the calls itself.
- Honest channel analytics: if a channel isn't working, you'll be the first to know — with numbers, not six months later.
// FROM_REAL_PROJECTS: a 400,000+ company lead radar · an outreach pipeline with human approval · a voice calling agent
// FORMAT: pilot campaign → pipeline (from ₽300,000) → ongoing support
AI mentor: guiding your clients through your own methodology
// SOUND_FAMILIAR?
- Client support runs on people: more clients means more payroll.
- A "GPT chat" answers in generic phrases and dilutes your methodology — a premium product gets devalued.
- Your methodology can be "extracted" from the bot and stolen.
- Payments, access, reminders and private groups eat up your admins' time.
// HOW_I_SOLVE_IT
- An AI mentor guides each client through your proprietary methodology step by step: the program and rules are fixed in code, so the model can't drift into generic advice; ~200 paying members are in AI-guided programs for a current client.
- Methodology protection: attempts to "extract" the content are detected and blocked — zero leaks since the protection went live; branded phrases are quoted verbatim from an approved bank.
- Conversation economics: support costs cut 5–6× with no loss of quality — measured on real dialogues.
- A complete admin-free loop: payment → access → guidance → reminders in the client's time zone → renewal; if something crashes, the system restarts itself.
// FROM_REAL_PROJECTS: an online school's AI agents, ROI 420% · a B2C service with 10,000+ users
// FORMAT: pilot on one program → full loop (from ₽300,000) → ongoing support
I don't sell AI where a simple script would do — and I'll tell you so straight away.
// CASE_STUDIES
Automation cases: it works and it pays for itself
No client names — deliberately: anonymity here is a principle, not an excuse. Every case follows the formula: industry + scale + result.
ROI 385%A distributor's documents with no manual entry
The problem. A distributor: 3,000–5,000 source documents a month, accounting retyping each one into the ERP for 5–10 minutes, errors surfacing at month-end close.
The solution. A recognition pipeline: scan/photo/PDF → verified structured data → ERP. An anchor check that it's the right document, rejection of unreadable scans, disputed items routed to human review.
The result. 10–30 seconds per document instead of 5–10 minutes · 99%+ accuracy · 95% fewer manual-entry errors · ROI 385%.
ROI 420%AI agents for an online school
The problem. An online school with a proprietary methodology: client support was capped by headcount, and the methodology was leaking out through bot "extraction".
The solution. A family of AI agents: a mentor guides members through the program step by step (rules fixed in code), extraction protection, and a full loop of payment → access → guidance → renewal.
The result. ~200 paying members in AI-guided programs · zero methodology leaks since protection went live · conversation costs cut 5–6× · ROI 420%.
ROI 310%Document processing for a construction company
The problem. A construction company: certificates, delivery notes and invoices from dozens of subcontractors in different formats; manual entry couldn't keep up with the build.
The solution. A single intake pipeline with recognition and routing by project site, reconciliation against the contract registry, and export to the accounting system.
The result. The document flow handled with no extra headcount · a transparent history for every site · ROI 310%.
94% accuracyPurchase forecasting for a 15-store chain
The problem. A retail chain buying "by gut feeling" — either stockouts of bestsellers or cash tied up in an overstocked warehouse.
The solution. Demand forecasting on the chain's order history, automatic alerts to purchasing managers, accuracy monitored on every cycle.
The result. 94% forecast accuracy · purchasing without stockouts or overstock · managers work from the system's recommendations, not intuition.
13,000+An AI sales agent over a 13,000+ SKU catalog — with zero hallucinated prices
The problem. A wholesale equipment supplier: a 13,000+ SKU catalog, inbound inquiries in two channels, managers falling behind, a "button bot" annoying customers.
The solution. A natural AI consultant in both channels with a managed funnel: prices and specs only from the product feed, legally sensitive wording locked behind forbidden-phrase lists, hot leads handed to managers with context.
The result. Replies in seconds at any hour · zero invented prices over the entire run · stress-tested by 15 virtual shoppers across 16 scenarios before launch.
10,000+ usersA B2C bot service with paying subscribers
The problem. A mass-market bot service: growth to tens of thousands of users, live subscriptions, and a migration to a new architecture where downtime wasn't an option.
The solution. A rebuild for scale: support for 19 source platforms, live-database migration with no losses, self-healing after failures, a payment loop.
The result. 10,000+ users · paying subscribers · the migration lost not a single account — proof of handling real money at B2C scale.
Want the same result? Let's discuss your project.
// CLIENTS
Who I work with
There are no logos or names here — deliberately. Everything you've read above is anonymized so thoroughly that no client can be identified. Your data will be treated the same way.
//SCALE_01
Owners of small and mid-sized businesses
One specific pain: inquiries, documents, call monitoring. You start with a 1–2 week pilot — fast, measurable, no "six-month project".
//SCALE_02
Mid-sized companies, 50–500 employees
Operations, CRM, document flow, dashboards for management decisions. ERP and Bitrix24 integrations — handled with care, always with backups.
//SCALE_03
Groups with multiple legal entities
Portals with role-based access, multi-entity setups with no data crossover. My record to date: 4 ERP databases and 2 Bitrix24 instances at once.
// Key clients have been with me for years — with new tasks every week. The typical path: pilot → project → support by subscription → the next task. I work directly with the owner or the head of the business line. NDA on request. Geography: Russia + international projects.
// NOT_A_FIT_IF
- "Build us some AI so we have AI" — projects without a measurable goal.
- Projects with no access to the actual decision-maker.
- Tasks where a simple script would do — I'll say so at the free review and won't take your money.
// PROCESS
How a project runs
- 01
Project review
A 30-minute call, free, in plain business terms, no strings attached. An honest answer on whether you need AI here at all.
- 02
Pilot on a narrow slice
1–2 weeks, a fixed price, a measurable acceptance test on your data.
- 03
Launch
2–6 weeks. Every change goes through: backup → tests → verification on a live scenario. Non-trivial changes get an adversarial review before rollout.
- 04
Acceptance
"The numbers match the reference" — plus an honest report: what was verified and what wasn't.
- 05
Ongoing support
Monitoring, incidents, system evolution. It lives and grows together with the business.
// The thread running through the whole process: we don't move forward until you say "go".
// PRICING
Logic instead of a price list
STEP_1 // FREE
Project review
30 minutes. We find out whether there's real money in this for your business. If you don't need AI — I'll say so.
STEP_2 // FIXED PRICE
Pilot
1–2 weeks, a fixed price, a measurable acceptance test on your data. The pilot cost is credited toward the full project.
STEP_3 // FROM ₽300,000
Project + support
The price is quoted after the review — it depends on the task, not on a package. Ongoing support runs by subscription.
// AGENCY_TEAM_VS_ONE_ENGINEER
- 5 roles in one person: architect · backend · DevOps · security · data
- No markup for management, an office, or a "project team"
- One timeline — nobody waits on anybody
- One point of responsibility — the person who writes the code answers for the result
// ABOUT
The engineer behind these systems
Independent AI engineer. I take AI products all the way to production single-handedly — with real paying users and real money in the loop. 250+ production tasks in the last year, 72 of them written up as verified case studies — I'll show you the registry at the review, under NDA.
- I cover the architect, backend, operations, security and data roles myself — one person instead of five.
- Every production incident gets a root-cause analysis and a regression test — mistakes don't repeat.
- Engineering honesty: if an approach is a dead end, I'll say so straight, with numbers — instead of defending the hours already spent.
// 27 AI agents across 4 servers migrated to a new model in a single pass
// 10+ production servers under management
// 200+ regression tests on the flagship product
// A separate specialty — security: I get called in to investigate compromised servers; keys and access rotated with zero downtime
// Speaker at artificial intelligence conferences
When a project calls for more scale, I bring in trusted specialists — but you always work directly with me.
// APP_STORE
Products you can verify right now
My own apps, not client work. Each one can be opened on the App Store — the cards below are pulled from there automatically.
Need your own app or a subscription product? That's me too: from idea to publication, including taking over a project from a contractor who vanished. Let's talk →
// FAQ
Common questions
What if the AI lies to my customers?
Prices and terms come only from your database or product feed — the model simply has nowhere to "remember" wrong ones from. Legally sensitive wording is locked down with forbidden-phrase lists. Before launch, every bot is stress-tested by 15 virtual "customers" across 16 scenarios. The rule is simple: the model generates — the code guarantees.
Will my data end up in someone else's cloud?
By default everything runs on your own server. For restricted environments there is a fully local option with no external AI services at all — including speech recognition. The model can be swapped without stopping the service.
You work alone — what happens if you get sick or disappear?
The code and the servers belong to you. Documentation is written from day one, every system has a rollback plan, monitoring and self-healing. Any engineer can pick it up — proven in real project handovers.
I'm not technical — how will we communicate?
In plain business terms: pain → solution → numbers. No "embeddings" in the chat. Checkpoints at every stage: we don't move forward until you say "go".
How much does it cost, and why is there no price list?
The price depends on the task, not on a package: free review → fixed-price pilot (credited toward the project) → project from ₽300,000 → support by subscription. Timelines: pilot 1–2 weeks, launch 2–6 weeks.
What if it doesn't work?
That's exactly what the pilot is for: a measurable acceptance test on your data. No serious money is spent before the numbers are in. And you always get an honest report — what was verified and what wasn't.
Do you work with our ERP / amoCRM / Bitrix24?
Yes. My record to date is a group of companies with 4 ERP (1C) databases and 2 Bitrix24 instances at once. Also amoCRM, RetailCRM, telephony, email, and messengers (Telegram, WhatsApp, MAX).
Will you sign an NDA?
Yes, on request. Anonymity is my default principle: there is not a single client name on this website.
// READY_TO_START?
Let's discuss your project
Describe the task in your own words — no technical brief needed. I'll review it and tell you what AI can do here, what it costs, and where to start.
// Telegram: t.me/aixenixcom
// Phone: +7 901 745-22-88
// Email: info@aixenix.com
// Geography: fully remote across Russia (MSK time zone) and international projects