Hardparse
A vision-language model running entirely on your Mac. Tables, formulas, charts, seals, handwriting — parsed locally, no network round-trip, no data leaving the machine. Ships as a native app built on MLX, CoreML, and Metal.
$ hardparse ./invoice_2026_q1.pdf → loading model qwen2-vl-7b · mlx · 4-bit → page 1/3 0.82s → page 2/3 0.79s → page 3/3 0.85s { "invoice_no": "INV-2026-0412", "issued": "2026-04-02", "currency": "EUR", "line_items": [ { "sku": "A-104", "qty": 12, "price": 48.00 }, { "sku": "B-220", "qty": 3, "price": 910.50 } ], "total": 3307.50, "seal": "detected · bottom-right" } ✓ parsed 3 pages in 2.46s · 0 bytes sent over network
Full CV pipeline for detecting welds in pipeline inspection data. Found a position-dependent sensor bias nobody had noticed. Running on fuel pipelines across the country.
Raspberry Pi + Hailo NPU. Detects aircraft, locks a gimbal on them, measures range with a laser, fuses ADS-B, computes GPS. Real-time, edge-only.
30+ page platform with ABAC policy engine, hash-chain audit trails, mindmap document views. Built the compliance layer from scratch — off-the-shelf couldn't handle classification levels.
Person extraction across 128 Polish regulatory PDFs. Evaluated 46 LLMs through a unified pipeline, wrote preprocessing that made inference near-free. Picking the model was harder than running it.
On-device vision-language model for tables, formulas, charts. Fully offline.
hardparse.com↗ ML Benchmark Tracker CodeSOTAIndependent SOTA tracking across 17 ML research areas. 286+ verified results with implementations.
codesota.com↗ AI Recommendations TasteRayConversational recommendation engine. Builds a taste profile through dialogue, not collaborative filtering.
case study↗RAG solutions for enterprise clients. LLM pipeline monitoring with Langfuse. Model selection and retrieval optimization.
Signal and image processing for industrial non-destructive testing. Anomaly detection, binary sensor pipelines, visualization dashboards.
GenAI platform integrating LLMs into business workflows. FastAPI, LangChain, LlamaIndex. Production deployments.
Transaction systems, API integrations, async processing for a fintech platform.
Education
- Computer Science, PJATK (in progress)
- Physics, University of Warsaw
- XIV LO Staszica, Warszawa IYPT 2012 laureate
Certifications
- Google Cloud Associate Engineer
- Azure AI Engineer
I like systems where the model has to be good — where latency is real, the hardware is small, and the user will notice the instant the inference trips. Pipelines, aircraft, Macs. Voice belongs in that list. If you're shipping real-time models to real users, I'd like to talk.