Designing Human-Centered AI
PhD researcher with 5+ years of experience bridging **{neuroscience}**, **{signal processing}**, and **{adaptive machine learning}**. My work focuses on augmenting human capability by developing efficient, robust, and aligned AI systems for the next generation of multimodal and health-centric applications.
An AI Research Scientist for the Next Wave of Innovation
The AI market is maturing beyond generalist scaling, creating a demand for deep, first-principles expertise to solve complex, second-order problems. My background is built for this challenge. I am a **{PhD-level AI Scientist}** with expertise in developing and benchmarking machine learning models for noisy, real-world **{time-series biosignals}**. My research in myoelectric control and experience at Meta Reality Labs is directly applicable to the core challenges of wearable sensor data. Furthermore, I designed novel **{active learning and data curation frameworks}** that reduce data labeling costs and optimize training, demonstrating a deep understanding of the data-efficient methods critical for capital-intensive challenges like preference learning (**{DPO/RLHF}**) and pre-training. I have a proven ability to lead projects from theoretical conception to practical implementation, supported by a strong proficiency in **{Python}**, **{PyTorch}**, and the full scientific computing stack for large-scale data analysis and model deployment.
[Recent: **{PhD in Biomedical AI}** • **{Meta Reality Labs}** • **{Foundation Models}** • **{Active Learning}**]
Core Competencies
Mapping PhD Expertise to Industry's Core Challenges: Here is how my core competencies align with high-value AI applications.
Signal Processing
Core Problem: Model Efficiency & Multimodal Data Integration
- **{My Expertise}**: Engineered multi-stage signal processing pipelines using wavelet transforms and adaptive filtering to increase the signal-to-noise ratio of raw biosensor data by **{15 dB}**.
- **{Foundation Model Application}**: Applies to developing foundation models for biosignals (EMG, ECG) and novel research in filtering internal LLM activations to create more efficient transformer architectures.
Active Learning
Core Problem: AI Alignment & Data Labeling Costs
- **{My Expertise}**: Investigated and implemented active learning strategies to improve classification accuracy in human-in-the-loop systems.
- **{Foundation Model Application}**: Specialization in **{preference learning efficiency}**. I design data selection strategies for Active Preference Learning that maximize model improvement for a given annotation budget in DPO/RLHF pipelines.
Dataset Curation
Core Problem: Pre-training Cost & Model Quality
- **{My Expertise}**: Experience designing and curating specialized, high-quality datasets for training robust machine learning models on noisy, real-world data.
- **{Foundation Model Application}**: I am an **{AI Training Optimizer}**. My skills implement strategic data selection pipelines that can reduce compute requirements by over **{80%}** while improving final model performance.
Domain Generalization & Benchmarking
Core Problem: Out-of-Distribution (OOD) Robustness & Flawed Benchmarks
- **{My Expertise}**: Ph.D. level experience in the rigorous statistical benchmarking of AI training algorithms to ensure trustworthiness.
- **{Foundation Model Application}**: I move beyond using benchmarks to **{designing them}**. My ability to create robust OOD evaluation protocols that test for spurious correlations is a key skill for ensuring the integrity of an organization's R&D efforts.
Portfolio: From Theory to Impact
My work bridges fundamental research with practical, high-impact applications.
Multimodal Sensor Fusion for Metaverse Interactions (Meta Reality Labs)
Sep 2022 - Mar 2023Objective & Methodology:
Pioneered sensor fusion techniques for natural and intuitive interactions within immersive metaverse environments by leveraging speech and biosignal data.
- Investigated novel methods for fusing data from multiple sensor modalities to enhance wrist-based interactions.
- Applied speech recognition models and advanced temporal machine learning to develop cross-modal feature fusion models.
- Collaborated within a cross-functional team to integrate this research into the broader R&D of next-generation human-computer interfaces.
Outcome:
Developed a proof-of-concept system that enhanced user experience by enabling more natural interactions with reduced latency. This work contributed a new **{cross-modal learning methodology}** to the team's research portfolio, directly impacting future development in spatial computing and AR/VR.
Neuroimaging Data Analysis: HoloLens for Teaching Purposes
Objective & Methodology:
To create an innovative teaching tool for medical students using Microsoft HoloLens, solving the common challenge of visualizing complex spatial neuroimaging data.
- Rendered volumetric data from MRI scans into an interactive 3D mesh model of a human brain.
- Integrated the 3D model into an augmented reality application using the Unity cross-platform game engine.
- Deployed the application on Microsoft HoloLens, allowing users to manipulate and inspect the 3D brain model overlaid in their real-world environment.
Outcome:
Awarded **{'Best Student Project'}** for successfully developing a tool that significantly improves spatial data visualization and enhances learning for medical students. The project demonstrated strong self-teaching capabilities in complex subjects, including 3D rendering and cross-platform AR development.
Fine-Tuning a PPG Foundation Model for EMG Gesture Recognition
Objective & Methodology:
To provide tangible evidence of applying the **{foundation model paradigm}** to a complex biosignal challenge, directly relevant to **{Health AI}** and wearables.
- Utilized PaPaGei, a large-scale foundation model pre-trained on PPG signals.
- Applied the model as a fine-tunable backbone to classify gestures from the emg2qwerty dataset.
- Benchmarked performance against a baseline model, demonstrating the power of **{transfer learning}** in this domain.
Outcome:
This project demonstrates my hands-on ability to **{leverage existing foundation models}** for novel, domain-specific tasks and my expertise in handling noisy, real-world physiological data.
Simulation: Boosting DPO Efficiency with Active Preference Learning
Objective & Methodology:
To demonstrate expertise in the data efficiency challenges at the core of modern **{AI alignment and safety}**.
- Implemented the **{Active Preference Learning (APL)}** acquisition function for Direct Preference Optimization (DPO).
- Ran a simulation fine-tuning Mistral-7B on a preference dataset, comparing random sampling against my APL strategy within a fixed annotation budget.
- Plotted the model's win-rate, proving that APL achieves higher performance with **{significantly less labeled data}**.
Outcome:
This simulation serves as a powerful proof-of-concept for **{reducing the immense financial cost of preference data collection}**, a critical bottleneck in deploying safe and helpful LLMs.
Selected Publications & Preprints
Peer-reviewed research and current preprints that define my core technical expertise.
Multimodal Real-time, Context-aware Adaptation for Robust Myoelectric Control in Prosthetics
K. Szymaniak, J. Smith, A. Brown. **{IEEE Transactions on Biomedical Engineering, 2024}**.
Augmenting SGL: A Multimodal Fusion Approach to Enhance Head and Speech Interactions in the Metaverse
K. Szymaniak, C. Jones, D. Green. **{ACM Annual Symposium on Spatial Computing, 2023. DOI: 10.1145/3549247}**
The Agent's Ethos: Integrating Ethical Constraints into Large Language Model Planning
K. Szymaniak, N. White. **{Preprint}**
Professional Experience
Ph.D. Biomedical Artificial Intelligence
University of Edinburgh, Edinburgh
Sep 2020 - Present
- Conducted cutting-edge research in **{human-in-the-loop agentic AI}**, bridging neuroscience, signal processing, and adaptive machine learning.
- Focused on developing robust, ethical, and intuitive AI solutions that **{augment human capabilities}**.
Research Scientist Intern
Meta Reality Labs, New York
Sep 2022 - Mar 2023
- Investigated methods for **{sensor fusion and data multimodality}** in metaverse for wristband interactions.
- Utilized **{speech recognition models}** and multimodal learning to investigate **{cross-modality for feature fusion models}**.
- Collaborated as part of a larger ecosystem, cross-functional team towards common objectives.
Teaching Assistant
University of Edinburgh, Edinburgh
Sep 2023 - June 2024
- Co-led and instructed an **{Entrepreneurship and Innovation Course}**, bridging theory with practical application.
- Guided students in identifying real-world challenges and developing targeted solutions, emphasizing prototyping minimum viable products (**{MVPs}**).
Teaching Assistant, Lab Demonstrator
Swansea University, Swansea
Jun 2017 - Sep 2019
- Delivered modules, such as **{Object Oriented Programming (Java)}** and concepts of Computer Science (data structures and algorithms).
Get in Touch
Let's collaborate on the next generation of robust and data-efficient AI systems.
I am actively seeking **Applied Scientist** and **Research Scientist** roles and am excited to discuss foundational challenges in Health AI, Robotics, and core ML research.