n8n · AI agents · RAG & memory · production systems

Your team is already using AI. It just isn't doing any of the work.

Chat windows don't run a business. Orchestrated systems do. Eight years in automation, three of them shipping multi-agent systems on self-hosted n8n — retrieval, persistent memory, tool calling, queue workers and the error paths that keep them alive at 3am. 3,000+ automations in production.

Message on WhatsApp Book a 20-minute call Download the overview (PDF)

Free process audit first. If there's nothing worth automating, I'll tell you that.

Real processes, before and after Time per cycle
Cost reportingArchitecture practice · compare reports, produce the final document
Before
After
5h → 30mper cycle
Client document collectionLaw firm · chase, receive, classify, check, file
Before
After
Days → 1hto clear the queue
Qualifying inbound leadsInvestment firm · ask, score, route, follow up
Before
After
Sales time → 0agent handles first contact

Same headcount, same ad budget. The hours come back and the revenue follows them.

Built on, every week
n8n — self-hosted, queue mode Claude · Gemini · OpenAI PostgreSQL · pgvector · Supabase JavaScript Code nodes Docker · Redis · Traefik Custom MCP servers Make.com · Zapier Airtable · Salesforce · Monday
What actually gets built

Six systems, wired into what you already use.

Nothing gets replaced. Your CRM, your inbox, your WhatsApp, your drive — they stay. The AI goes in beside them and starts doing the parts nobody enjoys.

AI agents that talk to customers

Answer enquiries on WhatsApp, web chat or email around the clock. Ask the qualifying questions a salesperson would, collect what's needed, book the call, and hand over a scored record instead of a raw lead.

WhatsAppWeb chatVoice & emailMulti-language

A CRM that fills itself in

Records created and updated without anyone typing. Stage changes, reminders and tasks fire on their own, and ad spend is tied to closed deals — so you see cost per customer, not cost per click.

SalesforceHubSpotMondayPipedriveCustom

Document intelligence

Contracts, invoices, payslips, reports and scans read automatically: classified, key figures extracted, summarised and filed where they belong. A person approves before anything is committed — every time.

OCRExtractionComparisonHuman approval

A content machine

Ad creative, product cards, carousels, short video and posts produced on demand, in your brand and in several languages — then scheduled and published without a person moving files between tools.

Copy & creativeVideoSchedulingBrand rules

Market & competitor capture

Pull listings, prices, reviews, social posts and public company data on a schedule, clean it, and put it somewhere you can act on — a dashboard, a sheet, or an alert when something changes.

Web & socialScheduledDashboardsAlerts

The plumbing underneath

The unglamorous half that decides whether any of it survives contact with real users: retries, error paths, logging, approvals, access control, and someone on the phone when it breaks. That someone is me.

n8nAPIs & webhooksSelf-hostedMonitoring
Under the hood

The parts nobody shows you in the demo.

A workflow that works once in a screen recording and a system that runs unattended for a year are different engineering problems. This is the second one.

01

Multi-agent architecture

A router agent decides intent, a retrieval agent fetches only what's relevant, an analyst agent answers. Each has its own tools, its own system prompt and its own failure behaviour — instead of one giant prompt that breaks the moment a user goes off script.

tool calling · structured output · guardrails · human-in-the-loop
02

Memory and retrieval that hold up

Persistent conversation memory on Postgres so an agent remembers a client across weeks, not just within a session. RAG over your own documents with deliberate chunking, metadata filters and re-ranking — because naive vector search returns confident nonsense.

pgvector · hybrid search · chunking strategy · re-ranking · citations
03

n8n at production scale

Self-hosted in queue mode with dedicated workers and Redis, so a heavy job doesn't block the rest. Sub-workflows for anything reused, environment-separated credentials, versioned exports, and concurrency tuned to what the downstream APIs actually tolerate.

queue mode · workers · sub-workflows · Docker · Cloudflare
04

Where the nodes stop, code starts

JavaScript Code nodes for the transformations no node covers: paired-item tracking, batching, deduplication, PDF assembly, cost accounting. When even that isn't enough I write the service — custom MCP servers and small REST APIs the workflows call.

Code nodes · MCP servers · REST services · webhooks · OAuth
05

Reliability, not optimism

Idempotent webhooks with dedupe keys so a retry can't double-charge or double-file. Retries with backoff, dead-letter paths for what fails, structured logging, and alerts to a human — because silence is the worst failure mode an automation has.

idempotency · backoff · dead-letter · logging · alerting
06

Token cost as an engineering metric

Cost tracked per client, per agent and per generation, then driven down: routing cheap work to small models, caching what repeats, scoping context so you don't resend the whole history every turn. The difference between a demo and a margin.

model routing · caching · context scoping · per-tenant accounting
Example flows

Three systems, node by node.

Simplified, but this is the real shape of them. Full walkthrough of any of these on a call — the actual canvas, the actual executions.

Document intake with human approval

Legal · regulated

Clients send documents over WhatsApp in whatever format they have. The system classifies, extracts and files them — but a person approves before anything is committed, because the practice is regulated.

Webhook→ dedupe key→ media download→ OCR→ LLM classify + extract→ Postgres upsert→ review queue (human)→ Drive filing→ CRM status→ client reply

Days of chasing became a one-hour queue. Rejected items route to a drafts folder instead of vanishing, and every decision is stamped with who made it.

Lead qualification agent with memory

Real estate · investment

An agent talks to the enquiry before a salesperson does: asks what a good rep would ask, scores the answers, writes a structured record and keeps following up on a schedule.

Inbound message→ session lookup→ agent + Postgres memory→ RAG over offers→ structured score→ CRM record→ label routing→ timed follow-up→ handoff to human

First-contact sales time went to zero. The rep opens a scored record with the conversation attached, not a phone number and a guess.

A mini-SaaS running on n8n

Own product · paying users

My own creative-generation platform: a 206-step pipeline behind a web app. Users sign in, submit a brief and get finished creative back — with quotas, billing and per-generation cost tracked. n8n is not just the glue here, it is the product's backend.

App request→ auth + quota check→ LLM storyboard→ image / video gen→ per-language typography→ FFmpeg render→ CDN upload→ preview + edit loop→ delivery + billing

Queue mode with dedicated workers so one long render never blocks another tenant, and token cost is recorded against the account that spent it.

Built, delivered, still running

Systems that clients operate without me.

Live demos of any of these on the call — the actual screens, not slides.

A cleaning company that runs itself

2025–26

The customer books online, an AI agent collects the job details, assigns the right cleaner, sends reminders, handles changes and cancellations, and keeps the calendar and CRM in sync. Two customer-facing sites and 80+ automations behind it.

Built from zero, handed to their team, running daily.
n8n · Airtable · Google Calendar · WhatsApp API · cleaner-4u.co.il

Law firm: intake, documents, cost per case

2025–26

Three agents handle client intake and case questions in a regulated practice. A custom CRM sits behind them with a review screen where staff approve every AI-classified document before it's filed. Ad spend is connected to case outcomes.

Document collection: days of chasing → a one-hour queue.
Claude API · Next.js · PostgreSQL · n8n · Meta Marketing API

Automated reporting for an architecture practice

2025

A system that compares cost reports against each other and produces the finished report. Built with the practice's lead architect, so the output matches what they'd have written by hand — without the transcription errors that came with doing it by hand.

Five hours per cycle → thirty minutes.
n8n · Google Workspace · PostgreSQL

Creative production at volume

2025–now

My own product: a platform that turns one prompt into ad creative, product cards, carousels and short video, with an editing surface, per-language typography and a multi-language interface. Paying customers, cost per generation tracked and reduced.

Proof that the content machine is not a slide.
LLM, image and video APIs · FFmpeg · Next.js · CLONY.AI
How an engagement runs

Audit first. Build second. Handover always.

Engagements usually run three months to a year, remote, working alongside your own team.

01

Process audit

A short call, then I map where the hours actually go and what each of them costs you. You get a written list of what's worth automating, what isn't, and the expected return on each — whether or not you hire me.

Free · about a week
02

Build in the open

Highest-return item first, so something is live in weeks rather than at the end. You see it working on your own data, we adjust, then the next one. Everything is written down as it's built.

Fixed scope per module
03

Handover and support

Your team is trained on the system, not dependent on the person who built it. Documentation, access, and a support line afterwards. Systems I delivered a year ago still run their daily operations.

Ongoing if you want it
Why this year and not next

The gap is opening between companies that shipped and companies that piloted.

The work is compounding

Every month a competitor runs on automated intake is a month of data, of faster response times, and of a team spending its hours on customers instead of copying fields between tabs.

It costs a fraction of a hire

The systems above cost less than one junior salary and don't take holidays. The question isn't whether to spend — it's whether the spend goes into headcount doing manual work or into removing the manual work.

The tooling finally holds

Two years ago this was demos. Today models read documents reliably, agents hold a conversation, and the orchestration layer is stable enough to trust with production. That's a recent change.

Daniel Alisov
Who you'd be working with

I'm Daniel Alisov.

Eight years in automation, the last three building and deploying multi-agent AI systems that businesses run on daily. n8n is my primary orchestration layer — self-hosted, in queue mode, with JavaScript where the nodes run out — and I've shipped 3,000+ automations and 50+ complete systems on it.

I was a marketing director before this, which is why I start with what a process costs you rather than with which model to use. You get one person accountable for the whole thing — discovery, architecture, build, deployment, training and the support calls — not an account manager between you and whoever is actually building.

I've trained 50+ specialists and I run a 400+ member professional automation community. Explaining a system to someone non-technical is a large part of this job, and I like that part.

8 yrsmarketing & automation
3,000+automations in production
3 yrsshipping agent systems
50+systems delivered
1mini-SaaS built on n8n

Start with the audit. It costs you a call.

Tell me what your team does by hand that they shouldn't. I'll tell you whether it can be automated, roughly what it takes, and what it should return — before either of us talks about money.

WhatsApp · +7 775 087 9355 [email protected] linkedin.com/in/daniel-alisov

Your details go to me and nowhere else. No list, no sequence.

Daniel Alisov · AI implementation & automation · working with clients in the US, Europe and Israel Русская версия