Eugene
Surkov

AI Engineer · Production Systems & Agentic Automation
Lisbon, Portugal (EU) · UTC+0/+1
hello@esurkov.comesurkov.comTelegramDownload PDFRemote · Contract / Part-time
Eugene Surkov
How I work

AI-native development. I design, build and operate production systems through agentic development — Claude Code, extended with agent pipelines I built myself rather than used off the shelf. The work is architecture, specification, evaluation design, debugging and iteration. Four systems are live in production today, each built and owned end to end by one person.

In production
4 systems live, built end to end by one person
Unattended
Photo studio pipeline running over a year with no staff
On real capital
Forex expert advisors, a funding rate system and on-chain DEX execution, live now
Agents
8 specialised agents across an 8-phase research pipeline
What I Build
Agentic systems & orchestration
Multi-agent pipelines with defined inputs, structured outputs and critique loops grounded in objective evaluation rather than model self-assessment
Automation pipelines
Process automation with deliberate boundaries — what is worth automating has run unattended for over a year; the rest stays human by design
Rapid prototypes & MVPs
Idea to working system in days through agentic development, then iterated against real production usage
Internal tools & dashboards
Flask monitoring interfaces and operational tooling built for daily production use, not for demos
Data pipelines & integration
REST and WebSocket ingestion, cleaning, validation, reconciliation across sources, structured persistence, pandas analysis
Production trading systems
Live on forex (MQL5), Binance Futures and on-chain DEXes — where correctness is enforced by the P&L rather than by a test suite
Key Projects
On-Chain Trading System — Robinhood Chain
Python · asyncio · httpx · JSON-RPC + WebSocket · eth-abi / eth-account · Uniswap v4 · SQLite · systemd · Oracle Cloud
  • Chain-log ingestion from an EVM L2 — launchpad bonding-curve trades and Uniswap v4 swaps decoded straight from logs into a ~9 GB dataset, with block cursors that backfill any gap after a restart
  • Multi-source market data: JSON-RPC and WebSocket with a fallback node, Blockscout, GeckoTerminal, DexScreener and a third-party wallet-flow feed
  • Live on own capital: execution on the bonding curve and on Uniswap v4 via Universal Router + Permit2, with gas, daily and total loss caps, an expiring live switch and paper mode enforced at the client; Telegram trade and risk alerts
  • Research before money: hypotheses pre-registered and paper-tested before any live rule; the same research on Base found no edge, so nothing was deployed there
  • Ops companion: a self-hosted Hermes Agent (Nous Research) briefed on the system, used for quick status checks and as an outside auditor, giving a second opinion on decisions made in the Claude Code workflow
  • Built in two weeks (Sep 2026) entirely through Claude Code from written specs and decision logs — ~250 commits, 600+ offline tests replaying recorded RPC/API responses. My part is architecture, specification, review and operation; the code itself is agent-written
Multi-Agent Research System
Claude Code · custom agent pipelines · Python · 8-agent orchestration
  • 8 specialised agents across an 8-phase strategy discovery pipeline, each with defined inputs and structured output formats
  • Autonomous loop: hypothesis → codegen → backtest → critique → iteration, with critique grounded in an objective evaluation function
  • The critique step reads walk-forward backtest results rather than the model's own assessment — the objective reward signal most agentic pipelines lack
Snap It Space Automation Stack
Python · Flask · AI pipeline · CloudFlare · Framer
  • Capture → AI retouching → validation → cloud upload → personalised delivery link → scheduled GDPR deletion, unattended
  • Internal Flask dashboard for live monitoring of every stage of the pipeline
  • Strict output specifications — naming, formats, folder structure — enforced by automated verification at every processing stage
  • 300+ clients served with no employees; running in production for over a year
Funding Rate System
Python · asyncio · pandas · Binance Futures API
  • Continuous collection and signal generation across 100+ perpetual pairs, normalised into a queryable structure
  • Multi-venue ingestion with rate-limit handling and cross-source reconciliation
  • The first design hit a sub-millisecond latency wall; reframed the problem and rebuilt it to work within retail-grade infrastructure rather than buying past the constraint
MQL5 Expert Advisors — Forex
MQL5 · MetaTrader 5 · strategy tester · walk-forward analysis
  • Two advisors running live, developed through extensive strategy research and optimisation
  • Running on real capital — correctness is enforced by the P&L, not by a test suite
  • Live behaviour tracked against the model each was built from