AVAILABLE FOR NEW PROJECTS

// web3 · ai · backend automation

Everything fails.
Mine gets back up.

I design, build and operate production systems — on-chain trading platforms, automation bots, AI pipelines, and the backends and internal CRM behind them. They run unattended, so half the code is reconnects, retries and reconciliation.

TypeScript Next.js 16 Python Web3 / EVM Redis PostgreSQL Docker LLM APIs
paradox-lab/ops — failover replay simulated
09:41:02INFOws.feed connected · latency 41ms
09:41:05INFOcopy-engine trader fill detected 0x8f…c3
09:41:08OKcopy-engine order mirrored in 3.4s
09:52:17WARNrpc.primary timeout — failing over
09:52:18OKrpc.backup reconnected · 0 events lost
10:14:33INFOredis.hot 25k tokens today · p50 ~50ms
10:14:39OKreconcile books balanced · drift 0
--:--:--#built to get back up on its own
$ replay failover.log --speed humanreal ones are uglier
400+
users in production — one analytics bot
25k+
tokens / day on its hot path
5
chains integrated · incl. a non-EVM L1
24/7
unattended operation

about

Self-taught, production-tested.

I'm self-taught, and almost everything I know comes from building systems that had to work in production — then fixing the parts that only break once real users are hitting them. A fair share of those failures traced back to bugs I'd shipped myself; that's where the recovery habit comes from.

Before any of this I spent six years as a manual QA tester, which means I learned how software fails before I learned how to write it. That order turns out to matter in 2026: anyone can generate plausible code now, and the real work is verifying it — scope, error handling, whether the tests assert anything or just run green. 860 of my commits in the main product repo touch test files.

Eighteen months for everything below. The arithmetic looks wrong until you account for what changed: a model does a lot of the typing now. The reviewing is still mine, and that's what the six years bought.

Generation got cheap. Knowing what will fail didn't.

Trading platforms, bots, AI integrations, a CRM product — the range is wide, but the job is the same: software a business leans on daily, where somebody notices within minutes when it stops. I don't just ship these systems; I operate them, and that changes how you build them.

I'd be building this stuff anyway — the two open-source tools in the grid below started as weekend projects for my own use, before anyone asked for them.

quick facts
  • rolefull-stack / systems engineer
  • background6+ yrs QA → dev since 2025
  • focusweb3 · ai · automation
  • moderemote · B2B preferred
  • codeTypeScript · Python · Swift
  • status● available for work

selected work

Running in production

Systems I designed and shipped end to end. Where the code is private or client-owned, the card links to a sanitized architecture writeup; open-source projects link straight to the repo.

Flagship Web3 / Fintech TypeScript · Next.js PRODUCTION

Multi-Protocol Prediction-Market Aggregator

A production trading terminal that unifies multiple on-chain outcome-market venues — including Myriad, Limitless and Hyperliquid — into one UI, with a Polymarket-driven copy-trading engine, cross-chain deposits and portfolio aggregation across protocols. Built end-to-end in Next.js 16 / React 19 across four signing models (EOA, Account Kit AA, EIP-1271, agent-wallet). Client code is under NDA; a sanitized architecture writeup is linked below.

Architecture: Next.js UI feeds a unified trading core with four signing models, integrating Myriad, Limitless and Hyperliquid, with a Polymarket-driven copy-trading engine and cross-chain deposits across five chains
4
integrations · 3 protocols + copy-trading
5
chains incl. non-EVM L1
600+
commits · Mar–Jun 2026
8
security & correctness fixes
Copy-Trading Engine
WebSocket worker + FIFO PnL, ~3–5s copies vs 60s polling
Account Abstraction
Alchemy Account Kit, EIP-7702 Safe wallets
Cross-Chain Bridge
LiFi deposits, PancakeSwap V3 collateral swap
CLOB + AMM Trading
Limit orders, local binary-search sell quotes
Agent-Wallet Signing
Silent L1 order signing, AES-256-GCM keys
V2 Protocol Cutover
pUSD migration, recovered stranded funds
View architecture writeup
Product · SaaS FastAPI · React 19 MCP · AI-native Jun 2026 → ecompilot.pl ↗ PRODUCTION

EcomPilot CRM

Internal CRM + ops platform I designed, built and run for EcomPilot — an e-commerce services company selling on Allegro and other marketplaces. FastAPI + React 19 over PostgreSQL — sales pipeline, tasks, calendar, finance and a WeasyPrint offer generator — fed by seven integration sources (Meta Lead Ads, Fathom, Google Calendar, HubSpot, WhatsApp, email, wFirma invoicing). A 31-tool MCP server exposes the live app to Claude for read and write. In production, login-gated, source private.

Integration hub: Meta Lead Ads (webhook to auto-lead), Fathom (notes matched by attendee email), two-way Google Calendar, WhatsApp archive, email, wFirma invoices matched by NIP and a one-time HubSpot migration feed EcomPilot CRM; a 31-tool MCP server exposes it to Claude, with a RAG assistant grounded in leads, notes and finance
31
MCP tools · read + write
2,064
backend test functions
7
integration sources
27
REST API modules
MCP server for AI agents
31 read + write tools expose the whole CRM to Claude, plus a RAG chat assistant grounded in leads, notes and finance.
Seven-source integration hub
Meta Lead Ads → auto-lead, Fathom → notes matched by attendee email, two-way Google Calendar sync, a WhatsApp archive, email, wFirma invoices matched by tax id, and a full HubSpot API migration — orchestrated with n8n.
Finance + offer tooling
Invoices, expenses and cash-flow forecasting, plus a WeasyPrint PDF offer generator with per-marketplace pricing for Allegro and beyond.
Sales pipeline + calendar
Configurable stages, lead management, tasks and a month/week/day calendar with drag + resize; team and per-rep performance on top.
Tested and audited
2,064 backend test functions, 163 frontend test files, a 37-finding security audit with fixes shipped in three waves, and adversarial code and UX reviews.
Boring infrastructure
PostgreSQL with 54 Alembic migrations, Dockerized, self-hosted. Each integration reconnects and re-syncs on its own.
Read the architecture writeup
AI / RAGBun · TypeScript2026

Catalogue RAG Assistant

RAG pipeline: product catalogues (216 + 208 pages) through measured extraction (pdftotext, 8.2% to 0.09% garble, per-page OCR routing) into PostgreSQL + pgvector, consumed by a Mastra agent with grounded citations

An AI assistant for a Polish horticulture supplier, answering from its real product catalogues — 216 + 208 pages chunked into PostgreSQL + pgvector. Two-dev build; my side: an extraction pipeline picked by measurement (pdfjs garbled 8.2% of characters, poppler 0.09%), citations, chat and admin, and most of the React UI and its tests.

216 + 208 pages8.2% → 0.09%pgvector
Read the architecture writeup
FintechPython PROD

Crypto Trading Analytics Bot

Crypto Trading Analytics Bot — a real-time token alert, a 24h performance report with score distribution, and the Telegram settings panel with per-user filters and thresholds

Telegram bot for real-time crypto signal monitoring — per-user filters, multiplier tracking, subscriptions and Solana / NOWPayments billing. A 24/7 fintech system on a Redis hot path with async dual-write to PostgreSQL.

400+ users 25k+ tokens/day ~50ms
Documentation
AI AutomationPythonOpen source

JobSniper — AI Job Monitor

JobSniper in action — an 80% match alert with AI reasoning, score-distribution statistics over 209 offers, and the Telegram control panel

Scans five job boards, parses your CV, and uses GPT-4o-mini to score each offer 0–100% with reasoning. High matches trigger an instant Telegram alert. Full observability with Prometheus + Grafana and a circuit breaker around the LLM.

5 boards GPT-4o-mini 0–100% scoring
View code
Developer ToolSwiftMIT

Claude Monitor Bar

Claude Monitor Bar — usage ring, 5-hour and 7-day windows, per-model breakdown

Native macOS menu-bar app that surfaces Claude Code API rate limits in real time — 5h / 7d / Sonnet windows — with a color-coded progress ring, usage-history charts, plan recommendations and self-installing auto-updates.

macOS native Keychain OAuth auto-update
View code

stack

What I build with

pkg:fullstack

Type-safe apps end to end — App Router frontends and async services that handle thousands of operations in parallel without blocking.

Next.js 16React 19TypeScriptBunPythonFastAPI

pkg:web3

EVM smart wallets, account abstraction (EIP-7702), on-chain trading, cross-chain bridges and crypto payments.

wagmi / viemethersAccount KitSolanaPonder.sh

pkg:data

Dual-write architecture — Redis for the hot path, PostgreSQL for persistence and analytics, typed migrations with Drizzle.

RedisPostgreSQLpgvectorDrizzle ORMAlembicWebSocket

pkg:ai & infra

LLM integrations with circuit breakers, containerized deploys, monitoring, graceful shutdown and auto-recovery — and verification of AI-generated code before it ships: scope control, error paths, tests that actually assert something.

LLM APIsMCPDockerGitHub ActionsSentryGrafana
how i work

Engineering principles

Transparency before sales

If your idea doesn't make technical sense or there's a simpler path, I'll tell you. I'd rather lose a contract than watch a client burn budget on something that won't work.

Understand, then build

Before I write code I need to understand what can break — what happens when the server crashes, the API errors, the data is incomplete. I ask before deployment, not after a client call.

Operate what you ship

I run the systems I build. When something breaks in production, the alert reaches me, not a backlog — and the fix usually ships the same day.

System, not script

A script needs you to run it. A system runs for you — with monitoring, logs, alerts and documentation. I build the latter.

contact

Got a problem to solve?

You don't need a finished spec. Describe the problem — we'll talk about whether automation makes sense, what it might cost, and how long it would realistically take.

If it's outside what I do well, I'll say so on the first call and point you somewhere better.

By submitting this form you agree to the processing of your data for the purpose of responding. Your data will not be shared with third parties.