Jacek Szybicki

AI-automated technical research

I build AI tools for technical research and engineering design.

In my own projects I automate technical research with AI: from literature reviews with source checking to design in CAD.

See the projects LinkedIn
01

Projects

1.1

Multi-agent system for technical research

Goal
Answers to technical questions built only from findings that carry a quote from their source.
How it works
A question is split into cells. In each cell agents set out several competing hypotheses and plan tests, about a third of which look for evidence against them. Researchers read papers and documentation, a sceptic raises alternative explanations, and code computes each hypothesis’s confidence from independent sources.
Verification
A finding stays only if its quote is on the page that was read. A fine-tuned Qwen3.5-4B checks this (LoRA trained on 1,700 Claude judgments): 0.929 agreement with Claude, against 0.674 for the previous 27B verifier.
Stack
Open Qwen3.8-27B model served with vLLM on rented GPUs; sources from arXiv, OpenAlex, Crossref, Europe PMC and the web through a private search engine. Semantic Scholar is integrated and waits for an API key.
  • 45technical studies
  • 76,000+findings with a source quote
  • 0.929verifier agreement with Claude
  • $1–3GPU cost of a typical study
Download the project overview PDF
One finding from the license plate recognition study

Finding

The character recognition model achieves 100% LP-RR with a minimum character size of 8×16 pixels, and the recognition rate drops below 36.4% when characters are smaller than 5×10 pixels.

Verbatim quote

“the character recognition model shows a 100% LP-RR with a minimum character size of 8 × 16 pixels. Then the recognition rate dramatically decreases under 36.4% when the characters feature a spatial size smaller than 5 × 10 pixels”

1.2

License plate recognition on low-cost embedded hardware

Goal
99% plate reads with ordinary CCTV cameras and a single-board computer, without a GPU.
Result
An engineering design built on 16,800+ verified findings from 1,400+ sources, including peer-reviewed papers and preprints.
Covers
The model pipeline, the pixels needed per character, camera placement and exposure, an error budget to 99%, and a training loop on site data.
  • 7stages of image analysis
  • 99.003%survival product of the loss budget
  • 3,000passes in the prototype acceptance test
  • 3.7 hon 2 GPUs, USD 13.08 rental
Download the case study PDF

1.3

CAD agent for FreeCAD

Goal
The AI agent works like a design engineer: from natural-language prompts it designs parts and complete assemblies in FreeCAD, from concept to production-ready documentation.
Scope
It models sheet-metal parts with flat patterns for laser cutting, 3D solids and parts for 3D printing. It builds assemblies with moving parts such as doors, flaps and drawers, and checks for collisions across the whole range of motion. It selects joints and uses ISO/DIN standard parts and catalogue components from a STEP library.
Engineering
It calculates strength (including FEM), thermal behaviour, tolerance chains and IP sealing, and takes manufacturing technology and cost into account. Production drawings follow ISO standards.
How it works
Prompts become model scripts for FreeCAD 1.1 running without the graphical interface, and each version is published as FCStd, STEP and DXF. A prompt can be pinned to a specific face of a part in the 3D viewer. Larger tasks are split between agents: designer, draughtsman, checker and knowledge researcher.
Knowledge
A knowledge base prepared by the research system, including 563 rule cards and 8,467 findings with quotes on sheet-metal design (bending, joining, IP sealing, thermal), 444 press brake punch and die profiles, and a separate base for 3D printing.
  • 123published model versions
  • 563design rule cards
  • 8,467findings in the knowledge base
  • 444press brake tool profiles
Download the project overview PDF
02

About

Since 2016 I have worked with devices and technical systems, from electrical installations and production management to product development for parking systems. I lead projects from the choice of technology and components, through work with suppliers, to production and deployment at the client. The other part of my work is AI in real company processes. I design the concepts and functional solutions and implement them with AI coding agents.

03

Contact

research@zaplab.sbs

Jacek SzybickiLinkedIn