Summer school6–17 Jul 2026
Tsinghua University, Shenzhen
Heading to Shenzhen for a two-week summer school at Tsinghua University (SIGS).
Summer school6–17 Jul 2026
NeurIPS preprintMay 2026
FTerViT: Fully Ternary Vision Transformer
New preprint on fully ternary vision transformers. Each weight is one of three values, which shrinks the model and cuts its memory-bandwidth needs, the real bottleneck on edge chips.
arXiv:2605.21171 ↗
HackathonForgis
Forgis Physical AI Hackathon
Competing in Forgis's physical-AI robotics hackathon, run with Google DeepMind and IBM.
▶ Watch the demo ↗
Doctoral exam2026
Passed my doctoral exam
Cleared the ETH doctoral exam, examined by Prof. Luca Benini, Prof. Michele Magno, and Dr. Nadim Maamari.
TalktinyML
Presenting at tinyML, Lausanne
Giving a talk at the tinyML event hosted at the Logitech HQ in Lausanne.
Research tripSep 2025
TSMC & AI hardware, Taiwan
Visited TSMC's Hsinchu fab and met semiconductor researchers across Taiwan.
Read the post →01 /
About
(about)
I run neural networks on a datacenter $10 chip.
I'm Szymon Ruciński, a PhD researcher in electrical engineering at ETH Zürich (D-ITET) and CSEM, part of SwissChips, the initiative strengthening Switzerland's semiconductor base. I run modern AI on small, power-constrained hardware — microcontrollers, AI camera sensors, NVIDIA Jetson — mostly for robotics and industrial perception.
In practice that means model compression (quantization, pruning), hardware-aware training, and perception pipelines sized to the chip they run on. Before the PhD I built production ML at Accenture, TransPerfect, Samsung R&D, and Visium: agentic AI for enterprises, custom translation models, computer-vision pipelines.
On the side I train open-source Polish language models and mentor students in ETH's Robotics and Analytics clubs. I've won AI hackathons for the United Nations and with Samsung's translation team. What I care about most: getting research out of the notebook and onto real hardware.
02 /
Experience
2025–now
PhD Researcher
ETH Zürich (D-ITET) & CSEM · Zürich / Neuchâtel
Model compression for transformers (quantization, pruning) and hardware-aware perception pipelines for robotics. I deploy on microcontrollers, AI camera sensors, NVIDIA Jetson, and DGX Spark, and train at scale on HPC clusters.
2023–2026
Machine Learning Engineer
TransPerfect (formerly Apostroph AG) · Zürich
Set up an in-house data center. Built custom LLMs that improved translation quality in French, English, German, and Italian, made the production model 5× faster to serve, and added retrieval-augmented search over the internal knowledge base.
Dec 24–Mar 25
Machine Learning Architect
Accenture Schweiz · Zürich
Built multimodal and agentic AI systems and MCP servers for enterprise clients. Cut call-center ticket resolution from six hours to five minutes at 84% success, on-premise and on Azure. Led a small engineering team across Berlin and Stuttgart.
Sep 22–Feb 23
Junior Machine Learning Engineer
Visium SA · Zürich / Lausanne
Built unsupervised computer-vision anomaly detection that improved faulty-sample identification by 97%. Found and fixed errors in a chemical manufacturer's dataset, which raised output by 20%. Also shipped a document data-extraction system and an internal package for prototyping CNNs.
2020–2021
Machine Learning Engineer
Samsung Research & Development · Warsaw
AI translation for low-resource languages, including Indian dialects. Built synthetic text-generation and quality pipelines over hundreds of terabytes, and shipped NLP features in Bixby. My team, SRPOL, won WAT-2021.
Worked with & for
03 /
Publications & Preprints
PreprintMay 2026
FTerViT: Fully Ternary Vision Transformer
PreprintFeb 2024
Efficient Language Adaptive Pre-training: Extending SOTA LLMs for Polish
Open model · LLM
krakowiak-7b
One of the largest open Polish LLMs when it was released. I cleaned and assembled the training data, fine-tuned the model, served it on cheap CPU-only hardware, and released it free for anyone to use.
Open model · Speech
Skryba-v1
A 1.5-billion-parameter Polish speech-recognition model that captions in real time. Fine-tuned on 50,000 audio and text pairs, and open-sourced on Hugging Face.
05 /
Community & Mentoring
Robot Learning2026–now
ETH Robotics Club
I'm a member of the Robot Learning Division, working on world models and reinforcement learning for robots. We compete in robotics hackathons with partners including OpenAI, Tesla, ABB Robotics, NVIDIA, Hugging Face, and AWS.
ETH Analytics Club2023–now
Hack4Good · Team Mentor
I mentor ETH student teams. One project built an NLP tool for WWF Switzerland, where I helped with the modelling and kept the work organized.
06 /
Education
2025–now
PhD · Electrical Engineering & Information Technology
ETH Zürich (D-ITET) & CSEM · Zürich / Neuchâtel
Supervised by Prof. Michele Magno & Dr. Nadim Maamari.
2021–2022
MSc · Data Science (Exchange)
Fachhochschule Nordwestschweiz (FHNW) · Brugg / Zürich
Supervised by Prof. Dr. Daniel Perruchoud.
2021–2022
MSc · Software Engineering & Machine Learning
Lodz University of Technology · Łódź, Poland
2017–2021
BSc · Information Technology
Lodz University of Technology, International Faculty of Engineering · Łódź
2020
BSc · Software Engineering (Exchange)
Fachhochschule Vorarlberg (FHV) · Dornbirn, Austria
07 /
Certifications & Bootcamps
- English C1
- German B2 (Goethe)
- Mandarin Chinese beginner (UZH, 1 semester)
- Model Parallelism: Large Neural Networks · NVIDIA DLI
- Transformer-Based NLP Applications · NVIDIA DLI
- AI: Modern Methods · Goethe University Frankfurt
- Building LLM-Powered Apps · Weights & Biases