Egor Suslin

Senior AI Engineer in Prague

I build AI systems and take them all the way to production, through the security and AI assessments that regulated companies require.

Connect on LinkedIn Download CV

I design, build and ship AI systems end to end: architecture, code, security, DevOps, and the security and AI assessments that insurers, banks, pharma and energy companies ask for before anything goes live.

What I build

Speech
Speech-to-text, text-to-speech and speech-to-speech systems, including a local voice assistant that routes what you say through local models before calling a cloud planner.
Whisper, Ollama, MCP, Rust
Documents and vision
Contract document analysis with NLP and vision-language models, and extraction pipelines that read structured data first and call an LLM only when they have to.
VLMs, NLP, structured outputs
Agents and RAG
Retrieval-augmented assistants and agent workflows with tools, human escalation, evaluation, and a governance layer for agents that spend money.
LangGraph, RAG, evals, MCP
Security and AI security
Security and AI assessments of applications, AI security, and the access control, secrets management and threat checks that get an AI application approved.
Banking, insurance, pharma, energy
Infrastructure and DevOps
The infrastructure under AI applications: containers, CI/CD, cloud and self-hosted servers, and insurance support automations running in production.
Docker, CI/CD, nginx, Azure, Google Cloud, DigitalOcean

Case studies

A local-first voice assistant that drives a Mac

Goal: Control the Mac by voice, keep everyday requests on the machine, and send only multi-step work to a cloud planner.
How it works: Whisper runs locally for speech-to-text. A small local model routes single-tool requests straight to tools; multi-step requests go to a Claude Code planner that calls the same tools over MCP. Commands only run inside allowed folders, risky tools ask for confirmation after their arguments are validated, and email is only ever drafted, never sent.
Evidence: A routing evaluation set measures accuracy and latency, and the tools are also exposed as an MCP server.

Rust, whisper-rs, Ollama, Claude Code, MCP

A control plane for AI agents that spend money

Goal: Let autonomous agents pay for things while the user's mandate is enforced before every payment, on whichever payment rail the agent uses.
How it works: Each agent carries a verifiable identity. A policy engine checks budgets, allow and block lists and categories before anything executes, and payments over a threshold wait for a person to approve them. A client-side guard for MCP, LangChain and the Vercel AI SDK fails fast before a payment outside the mandate. One audit ledger covers card and crypto, with an EU AI Act export for deployers.
Evidence: Live payment rails for card mandates, Visa TAP, Stripe, USDC over x402 and Bitcoin Lightning, TypeScript and Python SDKs, an MCP server, and a test suite in CI.

Next.js, TypeScript, Prisma, Supabase, viem

Verified creator profiles with privacy-first analytics

Goal: One verified profile that brings together a creator's social accounts and trading-app referral links, with analytics on top.
How it works: Social accounts are verified with OAuth and PKCE, wallets with Sign-In with Ethereum or Solana. Sessions store only a SHA-256 hash, OAuth tokens are encrypted with AES-256-GCM, and visitor IPs are kept only as a daily-rotating HMAC. Referral links are locked to each app's real domain. A stats worker takes Postgres snapshots under an advisory lock, and an AI profile writer is paid for on-chain, checked by a pure, unit-tested payment function.
Evidence: CI runs type checks, lint, a dependency audit and unit tests, then builds, migrates and runs Playwright end-to-end tests on WebKit, Android and desktop.

Next.js, React, Postgres, Drizzle, viem, Playwright, OpenRouter

A non-custodial wallet where no server ever sees the keys

Goal: Hold, send, swap and get alerts on two chains without any server ever holding the user's keys.
How it works: Keys come from a seed phrase and are encrypted on the device with Argon2id and AES-GCM. One core library powers a web app, a browser extension and a Telegram Mini App. The extension bridges web pages to a background worker that asks the user to approve every request; the Mini App's bot verifies Telegram's HMAC signature and freshness, with per-user rate limits. A strict CSP and auto-lock round it off.
Evidence: About 185 tests across 14 test files, a written key-rotation runbook, and a Docker deployment.

TypeScript, React, WXT, Kaspa WASM SDK, polkadot.js, noble crypto, grammY, Fastify

Property auction analysis, from catalogue to scored deals

Goal: Replace a manual spreadsheet workflow for analysing UK property auction lots.
How it works: A Playwright browser pool renders listings and walks catalogue pagination. Each page goes through an extraction chain: JSON-LD first, then OpenGraph, then links, and an LLM only for what is still missing, with the model switchable per run. Deals are scored and auto-tagged (below market value, yield, refurb, watch, skip) and stored in Postgres.
Evidence: The LLM is called only when cheaper structured extraction comes up empty, which keeps cost down.

Node, Playwright, Express, Claude via OpenRouter, Supabase

A booking assistant for a service business

Goal: Answer customer questions and take bookings at any hour, and hand over to a person when the assistant should not decide.
How it works: A LangGraph agent answers from the business's own content through vector search, checks availability and books through the booking system, takes payment with Stripe, and escalates to a human when needed.
Evidence: Customer-facing in production, with a test suite around the agent's tools.

React, Express, LangGraph, OpenAI, Claude, Stripe, Supabase

A market signals bot in Rust with user-held exchange keys

Goal: Real-time alerts across 11 signal types on the top 500 coins, with optional one-tap execution using each user's own exchange keys.
How it works: Market data is polled into rolling history; RSI, MACD and Bollinger bands feed a signal engine with rank tiers and cooldowns, which alerts on Telegram. If a user has connected keys, a tap places the order on Bybit, signed with hand-written HMAC-SHA256. Keys arrive by direct message, the message is deleted, and they are stored encrypted with ChaCha20-Poly1305. It starts on testnet by default.
Evidence: 50 unit tests across 7 modules and a multi-stage Docker build.

Rust, tokio, teloxide, reqwest with rustls, chacha20poly1305

Toolbox

Languages: Python, TypeScript, Go, Rust, SQL
Frameworks: FastAPI, React, Next.js, Astro, Express, LangGraph, LangChain
Models: Azure OpenAI, Vertex AI, Claude, Gemini, local models with Ollama, Whisper
AI engineering: RAG, agents, MCP, evals, NLP, vision-language models, speech-to-text, text-to-speech, speech-to-speech
Infrastructure: Docker, CI/CD, nginx, Azure, Google Cloud, DigitalOcean, PostgreSQL, MongoDB, Redis
Security: Security and AI assessments, AI security, access control, secrets management
Certificates: Azure Fundamentals (AZ-900), Azure AI Fundamentals (AI-900), Azure Data Fundamentals (DP-900), AI-103 in progress, Claude Code in Action (Anthropic, 2026)

Journey

2018 – 2022
Czech University of Life Sciences Prague
Bachelor's degree in Data Processing. From 2020 also a part-time data scientist, doing machine learning and data analysis in R for research projects.
2021 – 2024
PwC Czech Republic
Started as a financial data analyst, moved to junior data scientist, then data science associate, building Python tools, machine learning models and analytics.
2024 – 2025
Accenture
Senior data engineer on a generative AI project: data pipelines, LLM integration and backend services in Python.
2025 – 2026
MSD
AI/ML engineer in the Prague IT organisation, shipping machine learning and LLM solutions for the global business, with security review and compliance assessments.
2026 – now
NN
Senior AI engineer for an insurance group: AI applications from an empty repository to production, insurance support automation, contract document analysis, and the security and AI assessments behind each release.

Selected work

Under NDA I can name some companies and not others, but I can always describe the projects.

For clients and employers

AI applications for an insurance groupInsurancePython, Azure, CI/CD

Insurance support automation and contract document analysis, built from an empty repository and taken through security and AI assessments into production.

AI application for a bankBankingPython, security review

Built from scratch, then taken through the bank's security and risk assessments into production.

AI application for an energy companyEnergyDesign to DevOps

Design, build, security and DevOps of an AI application, through to production.

Agentic framework for an AI startupStartupAgents, LLM APIs

Senior engineer and AI consultant, helping design and build the company's framework for AI agents.

More of my own

News fact-check pipelineLLM pipelineGemini, Grok, LangChain

Pulls news from many feeds in parallel, removes duplicates, summarises and fact-checks with several models, and posts to Telegram.

AI game master for a tabletop RPGIn progressReact Three Fiber, Claude, TTS

A browser 3D tabletop game with a deterministic rules engine, AI party members and an AI game master with a voice.

Outside work

I build things for myself and the people around me, try new AI tools and other tech as they come out, and spend the rest of my time training or playing.

Training: Bodybuilding, climbing, calisthenics
Games: The Witcher 3, my favourite
Security: At work and for fun
Markets: Crypto and trading
Languages: English, Czech, Russian

Questions

Are you open to freelance work?

Yes: remote, hybrid or in Prague.

Can you take an AI application through security review?

Yes. I have taken applications through security, risk and AI assessments at a bank, an insurance group, a pharma company and an energy company.

Can you show client work?

Under NDA I can name some companies and not others, but I can always describe the projects and how I built them.

What stack do you use?

Python, TypeScript, Go and Rust, with the modern frameworks around them: FastAPI, React, Next.js, LangGraph. Models from Azure OpenAI, Vertex AI, Anthropic and Google, or local ones through Ollama. Docker and CI/CD on Azure, Google Cloud or my own servers.

Which languages do you work in?

English, Czech and Russian.

egor@prague:~$ connect

The fastest way to reach me is LinkedIn.

Connect on LinkedIn riharex420@gmail.com GitHub Download CV