Staff Front-end Engineer · Design Systems · Privacy-First AI Infrastructure
I ship production systems end-to-end — from design systems and front-end architecture to local-first AI infrastructure. 15+ years of professional experience, most recently as Front-end Tech Lead at CGI. Solo-built a subscription SaaS and a 4-tool privacy AI ecosystem.
Open to founding / staff engineering roles, technical co-founder opportunities, and selective consulting.
→ awolczuk [at] gmail [dot] com · LinkedIn · Warsaw · Remote
Architected a modern insurance PWA achieving a 96% Google Page Speed score:
- Component library with full Storybook documentation and testing
- Storyblok integration enabling visual editing independent of dev cycles
- AI-augmented delivery pipeline using Figma Code Connect to sync design tokens with production code
Full-stack subscription web app shipped to production. Stack highlights:
- Streaming LLM responses — Vercel AI SDK + Claude, token-by-token rendering
- Custom auth bridge — Clerk JWT decoded inside Supabase RLS policies (non-standard integration)
- Heavy compute in serverless — C library compiled to WASM, singleton-cached across requests
- Subscription billing — Stripe Checkout with dynamic pricing and promo codes
One philosophy across all repos: AI should work for you, not harvest your data. Local inference.
pseudonym-mcp — replaces PII with reversible tokens ([PERSON:1], [PESEL:1], [CREDIT_CARD:1]) before your prompt reaches the cloud, then restores original values in the response. GDPR-compliant pseudonymisation, fully offline, works with Claude, GPT-4, Gemini and any MCP-compatible client.
claude mcp add pseudonym-mcp -- npx pseudonym-mcpNode.js Apple Vision ecosystem for local OCR & image analysis on macOS — same engine, different surfaces:
- macos-vision — the foundation. Apple Vision OCR for Node.js, native and fast.
- macos-vision-mcp — MCP server for Claude Code & Claude Desktop. Pre-extracts text before your AI sees it, cutting token usage by ~97% on real documents.
- sortai — macOS CLI that OCRs every PDF and image in a folder with Apple Vision and writes native Finder tags (
#Faktura,#Umowa,#Bank…) and comments into file metadata — instantly searchable in Spotlight, no database. Offline via Ollama by default; cloud LLMs optional, withpseudonym-mcpmasking PII first.
claude mcp add macos-vision-mcp -- npx macos-vision-mcpsententim — deterministic MCP server that verifies whether a Polish court judgment signature actually exists before you cite it. LLMs hallucinate legal case numbers; this tool makes them stop. Zero LLM at runtime — purely SQLite + FTS5 lookups, sub-10ms, fully offline. Returns FOUND, NOT_FOUND, or AMBIGUOUS against a local database built from SAOS public data.
- Your Obsidian Vault as a Private Second Brain — Powered by Local AI — how to use
macos-vision-mcp+pseudonym-mcpto query your vault with cloud AI without exposing personal data. - Your Messaging Apps as a Private Document AI — Powered by OpenClaw — using
macos-vision-mcp+pseudonym-mcpinside OpenClaw to OCR and anonymise documents sent over WhatsApp, Telegram, or Slack before they reach any cloud LLM. - Apple Vision vs Tesseract — A 50-File OCR-to-Markdown Benchmark — head-to-head OCR comparison on 50 PDFs with identical LLM formatter input; Tesseract wins on CER, Apple Vision wins on structural quality.
- Privacy Tiers for Document AI — Three Pipeline Configurations — fully local vs local OCR + cloud LLM vs local OCR + pseudonymisation + cloud LLM: what each configuration protects, where the trade-offs are, and how to choose.
- LLM as a Bridge Between Qualitative and Quantitative Research — using LLM-as-coder and LLM-as-judge to classify interviews and project free-form text onto validated scales (PHQ-9, BDI-II, SF-36), with bias, non-determinism, and IRB caveats.
Outside software I serve on the management board of a large residential community in Warsaw, co-managing a 7-figure budget and vendor contracts.
