Artem Lisovtsov

Now

I co-founded Syntalith in Warsaw. We put AI to work in Polish companies, and I design each system and keep it running after launch: what it may do on its own, when it has to ask a person, and what happens when a client changes a price list or a vendor changes an API. I also teach people to work with Claude and ChatGPT, one to one or as a team, on their own work and without slides, and I run the AI-Native course: coding agents on your own repository, one to one or for a company.

One of the people I trained one to one is Kuba Koziej, co-founder of InterviewMe.pl and Zety, which BOLD acquired in 2019: a personal AI agent for him, and AI brought into his team’s daily work.

With a client it goes in four steps:

1. Call30 minutes with an engineer, free.
2. PrototypeA working system on your process and your data, also free.
3. ImplementationThe target and a fixed price go in the contract. 50% at signing.
4. AcceptanceThe other 50% once the target is met.

The target in the contract and the 50/50 split are for automations and custom tools.

Syntalith is an OpenAI Select Partner and a member of Anthropic’s Claude Partner Network.

Stack

Architecture

System architecture: event-driven services and microservices, consistency where the data is money, high load and scaling, moving off legacy systems. For AI, what a model may do alone, where a person signs off, and what has to stay in a private deployment.

Languages

Python, Go, Rust. PHP with Symfony and Laravel.

Models

Claude and OpenAI models, AWS Bedrock. Qwen and other open models under vLLM and llama.cpp, on our own MacBook and RTX 3090, or on a client’s servers when the data can’t leave the company.

Training

A detector of AI-written email: mmBERT with its upper blocks fine-tuned and calibrated to a set false alarm rate, with training data generated locally by Qwen 27B and SmolLM2. Classifiers on Jev. LoRA on Mistral and LLaMA, AWS SageMaker, SFT, DPO and GRPO pipelines. Polish text to speech with LoRA on VoxCPM2 and Chatterbox over 112 hours of speech, and a turn-taking model for voice agents.

Speech

ElevenLabs, Whisper, Amazon Transcribe and Polly, SIP telephony.

Documents

OCR on scans and photos, with a quality check before any model reads them. CAD drawings (.dwg) read through LibreDWG and ezdxf.

Search

OpenSearch, pgvector, embeddings from Titan and Cohere, RAG over internal documents. Knowledge bases that answer only from documents the person asking may open.

Agents

OpenAI Agents SDK, Claude Agent SDK, LangChain and LangGraph, PydanticAI, Vercel AI SDK, Hermes and OpenClaw. Tool calling, MCP servers over CRMs, ERPs, mailboxes and internal databases, structured outputs, planning and memory, multi-agent systems, prompt injection defences for agents that read mail and documents.

Evaluation

Traces in Langfuse. Test harnesses with simulated users, voice included, where must-pass scenarios are checked before averages. Models compared before release, and audits of what prompt caching saves.

Collection

Playwright and Camoufox workers with account leases and shared rate limits, for property portals and competitor monitoring.

Production

Chatbot APIs on AWS: Lambda, API Gateway, S3, IAM. Docker, Dokploy, VPS and Postgres. Logging, monitoring, latency, and model updates after launch.

This page

Plain HTML and CSS, and Three.js for the building.

Before language models

The backend and infrastructure I started with. Every AI system I build still stands on them.

Backend

PHP with Symfony and Laravel, Go with gRPC and GORM, Java, Rust.

Frontend

TypeScript, React, Vue, Angular.

Queues

Kafka, RabbitMQ.

Databases

MySQL, PostgreSQL, MongoDB, Redis.

Infrastructure

Docker, Vagrant, Kubernetes, Helm, Terraform, Ansible, Jenkins, FluxCD, AWS, GCP.

Monitoring

Prometheus, the ELK stack.

Experience

Working since 2019, with language models since 2022, long enough to see plenty of “agents” that were chatbots in new packaging.

Syntalith

Aug 2025 – now · WarsawCo-founder, AI systems architect. Dozens of systems since then. Email operations for a creator platform: every message is quarantined and screened for prompt injection, a draft goes out only if its body hash matches what a person approved, and 95% of replies now go out without anyone typing them. Browser workers that read OLX, Otodom, Gratka and Morizon and match new listings to clients’ briefs within the hour. A tender system that reads the CAD drawings and checks every line of the bill of quantities. Investment screening for funds, private models on clients’ servers, a knowledge base with permissions that leaked nothing in 750 test pairs, and sales pipelines and competitor price monitoring for the Polish market. Our own products are didntask.ai, which screens unwanted mail with models on our own servers, and a test bench for voice agents.

Andersen Lab

Jun 2022 – Apr 2026 · remoteSenior backend, then AI/ML engineer, then AI systems architect. High-load fintech and media systems, among them a forex broker and one of the largest TV broadcasters in the EU. Event-driven components and video processing first, later production AI: a RAG platform for internal documents, LLM agents with tool calling, semantic search and private model deployments, for hundreds of thousands of users a day.

Before language models

Andersen Lab

From Jun 2022 · remote · fintechGo and PHP behind a forex broker’s trading and partner systems. The main piece was an event-driven Go service (gRPC, GORM, Kafka, RabbitMQ) that kept client data in sync between two brands and the newer partner services: deposits, orders, transactions, financial scoring, rebates, withdrawals. With data like that the smallest inconsistency is a real difference in money, so every event was tracked, with Prometheus watching the system and ELK holding the logs.

Around it, a Go API pulling trade and account data out of MetaTrader 4 and 5 databases for reporting, tools that calculated partner rebates and PAMM accounts, and legacy PHP kept alive next to new Laravel and Symfony systems: a partner portal with referral statistics, rebates and reports, a partner CMS, and two CRMs, one of which opened a task for an operator whenever a new client registered.

Adapt

Jun 2022 – May 2023 · part-time, remoteBackend engineer. Rebuilt the backend of one of Lithuania’s largest electronics retailers, from a legacy system onto PHP 8.2, Symfony 5.4 and Laravel 9, working out the requirements with the client directly. I also reviewed code and planned the switch from the legacy API.

Open Mobi

Dec 2021 – Jun 2022 · WarsawBackend engineer on adQuery, an ad-tech platform on Symfony 5 and Laravel for creating, managing and serving rich media ads in real time under heavy traffic, and for storing and analysing the data that came back. I worked on its scaling, its architecture and its CMS. This is where I learnt what a system that has to hold up under big volumes looks like.

Mits

Oct 2019 – Dec 2021 · LublinFull-stack developer, my first job: custom CMS and CRM systems on Symfony, on Laravel and from scratch, online shops and APIs, integrations like TecDoc, and Vue and React when there was no backend work. The first place I saw how a business runs from the back office to sales.

Politechnika

Oct 2017 – Mar 2021 · LublinEngineer’s degree in computer science at Lublin University of Technology.

Human

I’m from Kyiv and I have lived in Poland since 2017, in Warsaw now. I built my first website in 2009, in plain HTML, and studied computer science at Lublin University of Technology, which I still call my beloved Politechnika. I speak Ukrainian, Russian, English and Polish, and I’m happy to talk in any of them.

My favourite moment at work is when a business owner stops talking about AI and starts describing their Monday: who answers the emails, where enquiries land, what gets lost. That is usually where the work is, and often something simpler than an agent is enough to start with. When that’s the case I say so on the first call, because I don’t like wasting people’s time.

On a computer with a wide screen there is a building behind this page. I wrote it in Three.js with one model, the hand, and almost no images, because a homepage should have at least one thing on it that nobody asked for. Choose “Walk the rooms” in the top right corner and go through it on foot: it starts in a room with one door, and through it is 1999. Every room after that is the next place I lived or worked, in order.

Contact

If a process eats half of someone’s day every week, write to me. Thirty minutes, then a written summary within two working days. If AI isn’t the answer, it says so.

Email: [email protected]
LinkedIn: linkedin.com/in/alisovtsov
Telegram: @syntalith
Syntalith: syntalith.ai, with the price list and a contact form