Projects#

EASE Health#

What: Amortised Bayesian inference (Bayesian neural network) for personalised migraine risk, with active-learning convergence in deployment. Trained offline on synthetic data; continuously optimises for information gain.
Origin: Gold on the Pfizer × Aava track at Junction Main.
Status: Accepted to Spark Finland’s 2026 autumn batch; building with Veikko Svynarenko and Aarni Konola.
Site: https://ease-health.org
Repos: ease-org · PlayerPlanet (EASE / Junction 2025)
Tags: Bayesian NN, active learning, N-of-1, health

What: Decision-support for tuning Deep Brain Stimulation parameters in Parkinson’s disease. Dense daily PDQ-39 data feeds a mechanistic “digital-shadow” forward model; amortised inference (SNPE) + Pareto multi-objective optimisation produce a recommended configuration the clinician reviews and applies. Human-in-the-loop by design, since regulation prohibits self-tuning brain implants.
My involvement: I worked on the original StimIQ with Atte Laakso, Nikolas Juhava and Niklas Keiski. The EuroTech × Hong Kong revision evolved without me — built by Atte Laakso, Marc Chen, Qilun Li, and Nikolas Juhava.
Status: Original advancing as a side project (Letters of Intent sent to Finnish DBS researchers); EuroTech × HK track continues independently.
Original preprint: StimIQ.pdf
Original demo: https://stim-iq.vercel.app
Original repo: PlayerPlanet/StimIQ
EuroTech × HK track (by others): Latt33/StimIQ · paper · video
Tags: SBI / SNPE, digital shadow, Pareto optimisation, clinical

What: Quantum Singular Value Transformation applied to credit and real-estate risk on Finnish data across 17 regions; deployed on the LUMI supercomputer and on IBM quantum hardware (Boston during the hackathon, Heron r3 for the scaling study).
Pipeline: Train NPE over different regimes → posterior samples for a factor-copula → block-encode the factor-copula parameters (single NPE forward pass) into a QSVT unitary → run QAE to estimate VaR + CVaR.
Honest result: ~10% VaR-95 tail-estimate error on a 12-asset portfolio on real IBM hardware; error scales with QSVT polynomial degree and circuit depth (peak ~11% at degree 20 on Heron r3).
Hackathon: Silver on the OP Pohjola track at Junction X OP Pohjola 2026.
Team: Milla Kolehmainen, Antti Nurminen, Silja Heiskanen, with mentor Juha Vesanto.
Companion paper: Implementing Credit Risk Analysis with Quantum Singular Value Transformation.
Write-up: Junction X OP Pohjola: QSVT4CRA — silver
Repo: PlayerPlanet/QSVT4CRA
Tags: quantum, HPC (LUMI), SBI / NPE, factor-copulas, QAE, QSVT

What: Depth-based policy model for the LeRobot SO-101 arm. Pose estimation that adds depth estimates from the raw feed as features to the network, substantially reducing training loss and train/val divergence versus the depth-free baseline.
Preprint: MANTIS-4D.pdf
Datasets: https://huggingface.co/datasets/k-kiirikki/top_wrist_so101_with_depth
Status: On hold; would love to validate on different setups and robustness.
Repo: PlayerPlanet/MANTIS-4D
Tags: robotics, physical AI, depth

What: Device localisation from advertisement-call signal strength, fitting physics-based non-linear models via least-squares.

What: Vision neural networks estimating coffee-pot fill level for the Physics Guild’s “kahvibot.”
Repos: PlayerPlanet/coffee_level_detection · PlayerPlanet/coffee_images

What: Browser-based card game (Cabo/Golf variant) — playable single-player vs 3 bots with special cards, throwing mechanics, and sound effects.
Link: Games
Status: Completed, playable.