Achievements

Rank 140 of 6,300 teams (top 2.2%)·Solo·Oct 2026

Healthcare

Challenge

Build a model that detects 12 kinds of abnormality on knee MRI, including ligament injuries, meniscal tears and fractures. Only a small share of the training scans were labelled; the rest came with the radiologists' written reports, from which labels could be derived.

Approach

Starting from a public notebook, Sekioka combined two lines of high-scoring publicly shared models with Sekioka's own models, weighting them differently for each type of abnormality. For Sekioka's own models, large language models read the reports and images to fill in missing training labels, and the image models included ones pretrained on medical images.

Rank 63 of 1,588 teams (top 4.0%)·Team of 2·Apr 2022

Environment

Challenge

Identify individual whales and dolphins from photos, using the shape and markings of their fins and bodies. The aim is to automate the photo matching that researchers do by eye.

Approach

In a two-person team with another Kaggle user, the team tried metric-learning models (ArcFace), which turn each image into a vector so that photos of the same animal end up close together, and pseudo-labelling, which feeds the model's own predictions back in as training data. The final submission was built on an ensemble that combines predictions from top publicly shared notebooks.

Rank 75 of 1,889 teams (top 4.0%)·Team of 4·Jun 2022

Intellectual property

Challenge

Score from 0 to 1 how close in meaning two phrases from patent documents are, within the context of the patent's technical field.

Approach

In a four-person team with other Kaggle users, the team started from a public baseline notebook, trained several pretrained language models, such as DeBERTa and BERT for Patents (a BERT model trained on patent text), and took a weighted average of their predictions.

Rank 163 of 3,633 teams (top 4.5%)·Solo·Aug 2021

Education

Challenge

Predict, as a number, how easy a passage is to read for students in grades 3 to 12.

Approach

Building on methods from public notebooks, Sekioka trained several versions of the pretrained language model RoBERTa, varying model size and training settings, and took a weighted average of three groups of models, tuning the weights over many submissions.

Rank 73 of 1,867 teams (top 3.9%)·Solo·Mar 2026

Life sciences

Challenge

Predict the 3D shape that an RNA molecule folds into, from its nucleotide sequence.

Approach

Sekioka started from a public notebook that combines two methods: one finds known structures with similar sequences and copies their shape, and the other is a deep-learning structure predictor (Protenix), used for sequences without a good template. Sekioka borrowed a technique from another high-scoring public notebook to split long sequences into pieces and stitch them back together, and also added support for molecules made of several chains.

Rank 165 of 2,605 teams (top 6.3%)·Team of 5·Nov 2021

Healthcare

Challenge

Predict how the airway pressure in the breathing circuit changes during a breath, from the ventilator's control settings (such as how far the air valve is open) and the properties of the lung. The data came from a ventilator connected to an artificial test lung.

Approach

In a five-person team with other Kaggle users, the team built several models, including neural networks that read the time series in both directions (bidirectional LSTM and GRU), combined them with a public notebook's predictions by taking the median, and also applied post-processing.

Rank 431 of 4,373 teams (top 9.9%)·Team of 4·Nov 2020

Drug discovery

Challenge

From data on how a drug changes gene activity and cell survival, predict the probability of each possible mechanism by which the drug works.

Approach

In a four-person team with other Kaggle users, the team built two neural networks on features prepared by normalising the data and compressing it with PCA, and averaged their predictions with a 7:3 weighting. The more heavily weighted one (70%), based on a public notebook, used transfer learning: it first learned from the scored labels together with extra labels that are not scored, then was fine-tuned on the scored labels alone.

Rank 161 of 10,148 teams (top 1.6%)·Solo·Oct 2026

Business simulation (farm)

Challenge

Build an AI agent, a program that decides its own actions, to run a virtual farm for a 30-day season. It chooses what to plant, how to care for its animals, when to buy land and when to trade on the market, and it wins by finishing with more money than its opponent.

Approach

Sekioka compared strong publicly shared agents, used their code as the base, and added improvements such as swapping a planned animal purchase or crop planting for a different kind to match demand from the town's shops and expected market prices. Each change was accepted or rejected by its win rate in automatically run matches under identical conditions, with each agent playing from both player seats, and AI agents carried out much of the work, from implementation to testing and submission.