RU EN

Dmitry Novoselov · AI agents and software development

AI agents
for business
you can trust with real work

I don't just build AI agents — I put them to work in your business processes. Behind my solutions are 25 years of engineering experience.

TURNKEY AI AGENTS

  • AI agent subscriptionfrom $300/month
  • AI employeefrom $1,100

CUSTOM DEVELOPMENT

  • SaaS & web platformsfrom $2,200
  • Mobile appsfrom $1,700
  • You see the work every week — on your own examples
  • The AI runs on your own subscription — no markup from me
  • An AI employee, its code and the data are yours
  • Contract and NDA before any work starts

ABOUT ME

Your agent is built by an engineer, not assembled in a no-code tool

For 25 years I've designed and built accounting systems, SaaS products and mobile apps. I build agents the same way: with access rights, checks, an action log and tests on real data.

For more than 15 years I worked as a lead developer at a wholesale company with its own manufacturing and a network of branches. I was responsible for the corporate ERP system that brought together the core business processes: purchasing, warehouse management, production, pricing and shipments.

So I start with the process, not the AI model: where an agent saves hours and where its mistake would cost money. Only then do I decide what to build in n8n, what to write as code — and where no agent is needed at all.

Dmitry Novoselov
25
years building software
50+
products running live

WHAT AGENTS DO

The agent takes over the routine that eats up the day

01

Leads and sales

Answers customers in a messenger from your price list, collects the request and passes it to a manager for approval. No lead gets lost at night or over the weekend.

Case study: a lead-handling agent in n8n →
Telegram amoCRM Bitrix24
02

Bookkeeping and cash

Categorises transactions, processes bank statements, receipts and payroll sheets, matches payments and sends a summary: how much cash there is and what needs attention.

Case study: an AI finance agent in a working company →
1C Banks Google Sheets
03

Customer support

Answers routine questions from your knowledge base and hands the hard ones to a person with a ready summary. It doesn't make up what isn't in the base.

Knowledge base Website chat Email
04

Documents and email

Processes incoming emails, invoices and contracts: extracts details, amounts and deadlines, files them in the right place and drafts replies.

Email PDF Excel
05

Meetings and tasks

Goes through call recordings, creates tasks with owners and deadlines, reminds about due dates and puts together a morning digest.

Call recordings Task tracker Telegram
06

Reports for the owner

Answers questions about the business in plain language: «why did margin drop in June», «who hasn't paid». Numbers come only from your systems.

Accounting CRM Spreadsheets

WHY YOU CAN TRUST THE AGENT

The concerns I hear most often

Each one is handled by how the agent is built, not by a promise. The agents in my portfolio work exactly this way.

  1. 01 «It will make mistakes» The agent proposes, you apply. It makes changes on its own only where its track record has already proven it accurate — and you set the threshold.
  2. 02 «It will leak our data» The agent has an employee's permissions, enforced by the server, not by the AI model. The most sensitive data can be hidden from the model or kept on a local model inside your perimeter.
  3. 03 «We won't know what it did» Every action of the agent goes into a log: what it read, what it proposed, who approved it.
  4. 04 «It will make things up or obey a stranger» The agent takes numbers only from your systems. Text from a customer, an email or a document is data to it, not a command.
  5. 05 «The AI bill will explode» The model runs on your own subscription or API key: you pay the provider directly with no markup from me, and usage is visible in a report.
  6. 06 «It won't fit our systems» I connect accounting software, CRMs, banks, Google Sheets, email and Telegram. Where no ready connector exists, I write one.

WHAT'S THE DIFFERENCE

AI agent or AI employee

An agent does a task. An employee owns an area of work — like a person on staff, except every step is logged.

AI agent
does a task
AI employee
owns an area of work
Scope
AI agentOne task or process: answer a customer, take a request, read a document
AI employeeA whole area, like a person on staff: finance, sales, purchasing
When it works
AI agentWhen someone asks or an event arrives
AI employeeEvery day on its own: scheduled summaries, checks, and it writes first when something goes wrong
Systems
AI agentOne or two connections for the task
AI employeeYour CRM, accounting, email and chats — with its own permissions, like an employee
Company knowledge
AI agentInstructions for the task
AI employeeThe company's rules and profile; it remembers past decisions
Works with
AI agentA customer or one person
AI employeeThe team: it knows who is writing, and everyone has their own permissions
Oversight
AI agentA log of every action
AI employeeA log, changes approved by a person, and autonomy that grows with proven accuracy
Price
AI agentby subscription from $300/month
AI employeefrom $1,100; the code and data are yours

WHAT IT COSTS

An agent by subscription or an employee of your own

An AI agent by subscription covers one task with no development fee. An AI employee is built for your company and stays yours. The exact price is fixed in the contract after the task is scoped.

ONE TASK · NO DEVELOPMENT FEE

AI agent subscription

from $300/month

An agent for one task: answers customers and takes requests, reads documents, prepares reports. No development fee — just a monthly fee, with a 3-month minimum term. The agent runs on my server, and keeping it working is my job.

  • One task, one or two connections
  • Server, support and updates are included
  • A monthly report on the agent's accuracy
  • The data stays yours

A WHOLE AREA IN YOUR SYSTEMS

AI employee

from $1,100

Owns an area of work like a staff member: finance, sales, purchasing. Every day it sends summaries on its own, notices what went wrong and prepares changes for approval. Built for your company — the code, settings and data are yours. The price depends on how many systems we connect and how much autonomy it gets.

  • Works in your CRM, accounting, email and chats with its own permissions
  • Knows the company's rules and remembers past decisions
  • Scheduled summaries and alerts when something goes wrong
  • A log of every action; autonomy grows with proven accuracy

Support for an AI employee — from $300/month. I monitor its work, add new skills and send a monthly report: how many proposals were accepted and where it was wrong.

PORTFOLIO

Agents at work

Agents I've built, and the products my engineering experience rests on. All designed and built by me, from the database schema to production.

Agents and AI in products

CASE 01 · AI AGENT
AI finance agent — inside a client company's books
The AI finance agent's change card with a confidence score, and the auto-apply settings (sample data)

Every day it categorises transactions, processes bank statements, receipts and payroll sheets and sends summaries to Telegram. It proposes changes with a confidence score and applies them on its own only where its track record has proven it accurate.

AI agent Bookkeeping Telegram Bitrix24
CASE 02 · AI AGENT · N8N
A lead-handling agent in n8n for an HVAC service
The agent core in n8n: an AI agent with a model, memory and three tools, answer checking and the model-failure branch

Answers an HVAC service's customers from the price list, collects the request and hands it to a manager for approval. An action log, refusing to invent prices and error handling — engineering practices on a no-code tool.

n8n Leads CRM Human approval
CASE 03 · AI AGENT · MARKETPLACES
An AI sales analyst for marketplaces
An ad cost ratio chart in the AI analyst's answer: ad spend on knives on WB went up while orders stayed flat

Answers questions about sales, sends a «what to do today» digest every morning and finds anomalies on its own: stock running out, a price rise that sank orders, ads getting pricier. Code does the maths; the agent only reads.

Hermes Agent Marketplaces MCP server Morning digest
CASE 04 · SAAS · AI ASSISTANT
FinRocket — management accounting and an AI assistant for small business
The owner’s home screen: cash on hand, its trend, revenue by product line and the cost structure Mobile app home screen: cash on hand and the trend for the period

Every bank account in one picture. The owner connects Claude or ChatGPT and asks about the finances in plain language: access follows the user's permissions, and changes happen only with their approval.

SaaS Finance AI assistant MCP server
CASE 05 · MOBILE · VOICE INPUT
A mobile app for the residents of a gated village
App home screen: the month’s charge, requests, passes and village news New request screen: choosing the recipient and topic, describing the problem, attaching photos Issuing a pass by voice: the app listens and parses a sentence with the guest’s details

Balance, invoices, photo requests and guest passes — from a phone. A pass is issued by voice: one sentence instead of six form fields.

Mobile Utilities Voice input Requests and passes

The engineering foundation

CASE 06 · SAAS
Legko i Prosto — accounting and analytics SaaS for 350 sellers
Sales dashboard: financial metrics, warehouse stock and order chart

A cloud management accounting and sales analytics service for marketplace sellers. 350+ clients, 500M+ transactions, the calculation core moved into the database itself.

SaaS Marketplaces Management accounting Sales analytics
CASE 07 · WEB PLATFORM
Mosaic — venture fund management platform
Investor and UBO registry: types, tax residency, share classes and KYC

SPVs, investors, cash movements and fund documents in one system. Carried interest over FIFO layers, event sourcing, KYC/UBO compliance. 76,000+ lines of TypeScript.

Web platform Fintech Venture fund Document workflow
CASE 08 · MOBILE
RouteBox — driver app and dispatcher web panel
Trip monitoring: vehicles on shift, routes on the map and a trip card with per-stop statuses Trip screen: the current stop, upcoming stops and closed ones with statuses

B2B delivery across Moscow: the day's trip for the driver, a live picture of every vehicle for the dispatcher. Offline mode for a full shift. Result: 31% less time per stop, 18% more stops per shift.

Mobile Logistics Offline mode Geolocation
CASE 09 · MOBILE · iOS
KShop — a mobile app for an Asian grocery store
KShop home screen: deal of the day with a countdown, weekly offers and bestsellers Product page: photo, discounted price, country of origin, weight, shelf life and customer reviews Recipe screen: the dish ingredients with prices and a button to add them all to the cart

An iOS app on top of the 1C-Bitrix catalogue: one-tap reorders, favourites that watch the price and recipes added to the cart with one button. Store turnover grew 23%.

Mobile iOS · SwiftUI E-commerce 1C-Bitrix

CUSTOM DEVELOPMENT

Whole products, when an agent isn't enough

01

Web Apps & SaaS

CRM and ERP systems, customer portals, marketplaces: from MVP to scalable platforms with billing, role-based access, integrations and analytics. From $2,200.

React Node.js PostgreSQL
02

Mobile Applications

Apps for iOS and Android: offline mode, maps and geolocation, push, camera capture and signatures. From the first build to release on the App Store and Google Play. From $1,700.

Flutter Swift Kotlin
03

Marketplace Accounting & Analytics

Management accounting for Wildberries and Ozon sellers: true profit, cash flow and per-product unit economics. Integration with marketplace APIs and reconciliation of reports that disagree with each other.

WB API Ozon API ETL
04

Audit & System Evolution

Inherited a product from a previous team and afraid to touch it? I'll work it out, put the architecture and the code in order, clear the technical debt and get the system ready to grow.

Legacy Performance CI/CD

HOW I WORK

From process review to an agent you trust

  1. 01 Process review I look at how the process works today and work out where an agent saves time and where a mistake would cost money. If plain automation without AI solves it cheaper, I'll say so.
  2. 02 Test set I collect 30–50 real examples with known correct answers. The agent is checked against them before launch and after every change.
  3. 03 Supervised launch I launch the agent on one process. Everything it wants to write is first seen by a person and approved with one click.
  4. 04 More autonomy Once the record shows the agent doesn't make mistakes on a type of change, we turn on auto-apply for it. Everything else still goes through you.
  5. 05 Support I monitor the agent, add new skills and send a monthly report on its accuracy. Or I hand the agent over to your team with documentation.

GUARANTEES

What protects you when you work with me

Four things worth agreeing on with any contractor before you start. With me they are written into the contract, not promised on a call.

You can walk away at any point

Servers, n8n workflows, code and settings of a purchased agent are registered to you from the start, and the data stays yours under any arrangement, including a subscription. If you decide to stop the project, you pay only for the work already done.

I bring in a team when it's needed

I have a pool of specialists I've worked with for years — design, QA, DevOps, backend. On large scopes I bring them in and pay them myself: your price and your contract don't change.

Contract and NDA before any work starts

I work officially, with proper closing documents. I sign the NDA before you show me your processes and data — I'll send my template or sign yours.

You see the work every week

Not a «60% complete» report, but an agent already answering on your examples. If something's heading the wrong way, it shows in week two rather than at handover.

QUESTIONS

What people usually ask before the first call

What if the agent makes a mistake?

Every worker makes mistakes; what matters is the cost of one. That's why the agent starts by only proposing: a person sees the change and applies it with one click or rejects it. Autonomy is switched on only where the record shows accuracy, and any automatic action can be undone.

Where does my data go?

The agent sees only what its role allows, like a new employee. Names, amounts and account details can be replaced with labels before they reach the model, and for the most sensitive data a local model can run on your server — then nothing leaves at all. The NDA is signed before we discuss specifics.

How much does an agent cost per month?

You pay the AI provider directly under your own subscription or key, with no markup from me. The AI finance agent in my portfolio runs on a regular ChatGPT Plus subscription. You don't even have to buy the agent: a subscription starts at $300 a month with no development fee, a 3-month minimum and support included. Model usage is visible in a report.

Why is this better than ChatGPT or a no-code builder?

ChatGPT answers, but it doesn't work inside your systems: it can't see your books, create leads or know your procedures. A no-code builder connects the systems, but permissions, approval, logging and quality checks are usually left out. I build in n8n where it's enough and write code where reliability is needed.

I don't know what to hand to an agent. Where do I start?

With whatever eats up your or your staff's time every day. On the first call we'll go through the process and I'll tell you honestly whether an agent is needed there. Sometimes plain automation without AI is enough — and then I'll say so.

Who don't you work with?

I don't promise to replace a department with AI, and I don't build a «chatbot in an evening» without access to your data — such a bot answers nicely and does nothing. I don't take on landing pages or sites on website builders. I don't work without a contract or on a «pay after launch» basis.

CONTACT

Tell me what work to hand to an agent

Describe in a couple of sentences the process that eats up time. I'll reply with my first questions and tell you whether an agent fits and at which level to start. If it isn't a fit for me, I'll say so straight away and point you to someone better suited where I can.

I reply within 3 hours on weekdays

I work under a contract and a signed NDA. Your data, code and business information are confidential from the first conversation, not from the first payment.