IT professional turned AI app builder. I build desktop applications powered by LLMs — things I actually want to use. After 10+ years in IT support I'm channeling that domain knowledge into building real tools with modern AI stacks.
Default stack:
Everything runs locally. No cloud, no subscriptions, no telemetry. The user owns their data.
🎯 Dataset Generator — flagship
No-code desktop app for generating high-quality synthetic datasets to fine-tune LLMs. Three-stage pipeline (topics → outlines → examples) with LLM Judge quality control, embedding-based deduplication, and HuggingFace Hub integration.
- Measurable results: +4.5pp on HumanEval, +5.0pp on HumanEval+ (Qwen2.5-Coder-7B-Instruct, 5-run average, error bars don't overlap)
- Plan-then-Execute engine — diverse, coherent datasets without repetitive patterns
- Real-time SSE dashboard, ShareGPT / Alpaca / ChatML export, HuggingFace Hub integration
- AGPL-3.0 licensed — open source with strong copyleft protection
Next.js 16 FastAPI Python SQLite OpenRouter Pywebview
Desktop app for tracking stocks, ETFs, gold, and Polish government bonds — with AI analysis powered by a Map-Reduce multi-agent architecture.
- Worker agents (Gemini Flash) analyse each position in parallel
- Manager agent (Gemini Pro) synthesises all reports into a single portfolio risk assessment
- Deterministic market regime detection (Panic Mode, Bond Shock, Oil Shock...) — no AI needed
- Full support for all 8 Polish MF bond types with live NBP/GUS CPI data
- Correlation matrix, benchmark comparison, Sharpe ratio, max drawdown
Electron React TypeScript SQLite OpenRouter Yahoo Finance Map-Reduce
I vibe code with Claude Code. That means I spec architecture, design data models, reason about edge cases — and let AI handle the heavy lifting on implementation. This lets me ship real, working desktop apps with complex features (multi-agent pipelines, native API integrations, proper local data persistence) without a traditional CS background.
I think this is a legitimate and underrated skill. The craft is in knowing what to build and why — not in typing faster.
Open to remote opportunities in AI tooling, LLM integration, and product development.
