┤ display ├
SKYVIEW

Andrzej Pomirski

AI/ML engineer and applied researcher based in Gdańsk, Poland, with over a decade of experience in speech technology and, more recently, agentic AI.

BASE
Gdańsk, Poland
FOCUS
agentic AI · speech technology
LINKS
githublinkedinorcidemailcv

Ask about his work

Instant answers come from a checked index of his profile. Every answer cites its section, like ┤#exp-amazon├; follow it to the source.

INSTANT ANSWERS ─── active

Fine-tuned Qwen3-4B, runs in your browser on WebGPU. Downloads once, about 1–2 minutes.

Panasonic

@ Remote (team in Palo Alto, CA)

Staff AI/ML EngineerForward-Deployed Engineer

Global Applied AI

Oct 2025 – Jul 2026

  • Built the core of KanpAI, Panasonic's enterprise agent platform: the SDK and orchestration engine behind every agent built on it. One of its first engineers and its second-largest contributor overall.
  • As a forward-deployed engineer, worked directly with other Panasonic operating companies, gathering requirements first-hand and iterating toward solutions for each business amid shifting, often ambiguous goals.
  • Built AIVA (see ┤#proj-aiva├).
  • Delivered agentic prototypes, several demoed to company executives: a manufacturing production-forecasting agent, and an on-site widget that predicts a visitor's likely path and guides them toward conversion in real time (see ┤#proj-purchase-intent├).
  • Contributed across KanpAI teams to speed up delivery: evaluation framework, CLI, low-code agent builder, developer tooling.

Staff AI/ML Engineer

Panasonic Well

Apr 2024 – Oct 2025

  • Primary developer and top contributor to Umi's AI layer (see ┤#proj-umi├).
  • Built the shared agent libraries beneath Umi and reusable service/agent templates adopted across the team.
  • Collaborated directly with Anthropic engineers during a strategic partnership.

Amazon

@ Gdańsk, Poland@ Cambridge, UK

Senior Software Development Engineer

Jun 2015 – Jun 2024

  • Promoted through four levels: intern (Jun 2015), SDE I (Sep 2017), SDE II (Oct 2018), Senior SDE (Jun 2022 – Jun 2024).
  • Built the core of a fully offline neural TTS system shipping in BMW and MINI cars (see ┤#proj-bmw-tts├).
  • Tech lead of the on-device TTS team (later Neural TTS on Device).
  • Developed and improved the on-device TTS engine powering Kindle e-readers, Amazon Lockers and Fire OS devices (Fire tablets, Echo Show), and delivered higher-quality voices for Alexa.
  • Created the initial Amazon Polly iOS SDK.
  • As a Lab126 intern in Cambridge, UK (Jul–Sep 2017), developed a network- and condition-aware test system for Amazon Cloud Cam.
  • Mentored engineers to promotion, helped candidates secure engineering roles, and advised master's students on three published papers through the Amazon–UPM/IPTC collaboration (2022–2023, see ┤#publications├).

Sii Polska

@ Gdańsk, Poland

Java Developer (internship)

Jul 2014 – Aug 2014

  • Developed parts of Phoenix, the product-lifecycle-management system Sii built for PUMA (ordering and tracking footwear-production tooling), using Java EE, Vaadin, EJB, Hibernate and IBM DB2.

AIVA in-flight assistantPanasonic Global Applied AI2025–2026

Panasonic's AI in-flight assistant for Panasonic Avionics Astrova seat-back units, built on KanpAI, with an avatar, voice, and layover planning. It was demoed to airline customers at Aircraft Interiors Expo.

Purchase-intent model and conversion widgetPanasonic Global Applied AI2025–2026

An on-site widget that predicts a visitor's likely path and guides them toward conversion in real time. It is driven by a gradient-boosted (LightGBM) purchase-intent model, trained end-to-end: about 0.95 AUC and about 80× lift on its top-ranked leads.

Umi family wellness coachPanasonic Well2024–2025

Panasonic Well's LLM-powered family wellness coach, since discontinued. Primary developer of its AI layer: built its multi-agent conversational AI from scratch to coordinate goals, scheduling and recommendations across a family, and took it through internal dogfooding and trusted-tester programmes with real families.

Offline neural TTS for BMW and MINIAmazon2020–2024

The core of a pioneering, fully offline neural text-to-speech system shipping in BMW and MINI cars (BMW Group news). Delivered in early 2024. Each brand has its own English voice under 150 MB, running at roughly twice real-time with under 200 ms to first audio on Pixel 4 XL-class hardware (paper).

LLM strategy for GreenWorlderConsultingApr–Jul 2023

Technical strategy and roadmap for an LLM-powered assistant on GreenWorlder, a Luxembourg-based sustainability social network: what the assistant should do, how to ground it in the site's own content, and an evaluation-first plan for choosing a model.

ChestShopOpen source2011–2016

An open-source Minecraft server plugin (Java) that lets players run shops from chests and signs. Created it, first releasing it as iConomyChestShop on the Bukkit forums on 12 February 2011, and was lead developer until 2016; community maintainers run it now. It has over 5.5 million downloads on BukkitDev and still ships releases (v3.13 in 2026).

Media-monitoring systemClient project2013–2015

A media-monitoring system for a power-grid construction company, which used it to follow press coverage of a high-voltage power line it was building. Written in Python (Flask, BeautifulSoup, Requests) and running on a Raspberry Pi, it tracked news articles for keywords about the project and presented the results through a small web service.

This portfolio assistantPersonal project2026

The assistant on this site runs entirely in the visitor's browser; no server sees the questions. Instant answers come from keyword search over curated Q&A; on desktops with WebGPU it can load an open LLM on request through WebLLM (Qwen3-4B), with this profile in context and a LoRA fine-tune for tone, refusals and citations (facts stay in the profile, not the weights). Built with Claude Code.

Research

Research interests: EEG-to-text decoding and brain–computer interfaces; neural speech synthesis; multimodal and representation learning; LLM evaluation; on-device and edge ML.

  • PhD: enrolled (2026) as a part-time doctoral student in the Doctoral Programme in Electronic Systems Engineering at Universidad Politécnica de Madrid (ETSI de Telecomunicación), supervised by Prof. Rubén San-Segundo.
  • Current work: EEG-to-text decoding; the paper is under review (see ┤#publications├).

Publications

Peer-reviewed:

  1. "Sign Language Motion Generation from Sign Characteristics." Gil-Martín et al., Sensors, 2023.
  2. "Sign Language Dataset for Automatic Motion Generation." Villa-Monedero et al., Journal of Imaging, 2023.
  3. "Exploring the Use of Contrastive Language-Image Pre-Training for Human Posture Classification: Insights from Yoga Pose Analysis." Dobrzycki et al., Mathematics, 2023.
  4. "Lightweight End-to-End Text-to-Speech Synthesis for Low-Resource On-Device Applications." Tura Vecino et al., ISCA Speech Synthesis Workshop (SSW), 2023.

Submitted, under review:

  1. "Contrastive Learning for EEG-to-Text Generation: A Vocabulary-Consistent Evaluation on the ZuCo Datasets." Pomirski, San-Segundo, Barra-Chicote. Neurocomputing, 2026.

Papers 1–3 come from the Amazon–UPM/IPTC collaboration (2022–2023), co-supervising master's students (methodology, validation and review on the sign-language papers; supervision and review on the yoga-pose paper). Paper 4 is from the on-device TTS work at Amazon: team lead, creator of the original lightweight model the paper builds on.

Skills

  • Programming: Python (strongest), C, C++, Java, TypeScript, Dart/Flutter, Bash, Perl.
  • AI/ML: LLMs and agents (OpenAI, Anthropic, Llama, LiteLLM), multi-agent systems, Pydantic AI, MCP, RAG, structured outputs; deep learning (PyTorch, TensorFlow); gradient-boosted models (LightGBM); Transformers, CNNs; text-to-speech.
  • Backend and infra: FastAPI, gRPC, PostgreSQL, Redis, Kafka, BigQuery, WebRTC, React, Linux, AWS, Docker, CI/CD, Datadog, Git.
  • Languages: Polish (native), English (C2, Cambridge Certificate of Proficiency, grade A, 2014); elementary German, Greek and Russian. Enjoys learning new languages.

Education

Universidad Politécnica de Madrid (ETSI de Telecomunicación), Madrid, Spain. Part-time PhD student, Doctoral Programme in Electronic Systems Engineering, since 2026 (see ┤#research├).

Gdańsk University of Technology, Gdańsk, Poland. Bachelor of Engineering (with distinction) and Master of Engineering in Computer Science, 2015–2020.

  • MEng thesis: "Deep Learning Methods for Spoken Language Identification" (2020).
  • BEng thesis: "Machine Learning for Road-Surface Quality Classification from Mobile-Device Accelerometers" (2019).

V Liceum Ogólnokształcące, Gdańsk, Poland. High school, mathematics, physics and IT class, 2012–2015.

  • Alongside high school: Akademia ETI at Gdańsk University of Technology (2012–2014).

Courses and certificates:

  • Machine Learning, Stanford University (Coursera).
  • Cambridge Certificate of Proficiency in English (CPE), C2, grade A (June 2014).
  • Basics of project and time management, S&O Partners (2015).

Awards and outreach

  • Gold Badge of the Graduate of Gdańsk University of Technology, for graduating with distinction (2019).
  • NASA International Space Apps Challenge, Global Award Nominee (2017): FireHelper, an app guiding people out of wildfire zones and alerting firefighters, from satellite fire data.
  • NASA International Space Apps Challenge, Local People's Choice Winner in Gdańsk (2018): Death clock, a simulation of how everyday choices affect global water shortages.
  • Invited by Amazon onto Artur Kurasiński's Technofobia podcast for Alexa's eighth birthday (December 2022), talking about the future of Alexa and AI (Amazon's post, in Polish).

Community and volunteering

  • Member of Hackerspace Trójmiasto, the Tricity hackerspace (2016–2019).
  • Member of the student research group "Gradient" at Gdańsk University of Technology, focused on machine learning and neural networks (2016–2020).
  • Member of the Faculty Student Council (WRS ETI, 2015–2016): represented students to the faculty authorities, took part in Faculty Council meetings and organised student events; elected in 2016 to the Faculty Electoral College as a student representative.

Interests

[gliding · SPL licence] [motorboating] [cars] [motorsport]

[3D printing] [electronics] [DIY] [smart home]

[amateur DJing · EDM]

[board games] [puzzles] [video games]

[travelling · Greek islands] [travel magnets]

[baking] [cooking] [craft beer] [coffee]

[languages] [psychology] [design]

Contact