I'm an ML engineer and data scientist at Fukushima Galilei (フクシマガリレイ株式会社). I build predictive maintenance for refrigeration equipment: models that warn us about a failure a few days before it happens. I trained as an electrical and mechanical engineer, so I start from how the machine works and add the machine learning after. On the side I build projects in search, wearables and simulation, and I write up what went wrong as well as what worked. I work in English and Japanese.
Fukushima Galilei (フクシマガリレイ株式会社)
I build predictive maintenance for commercial refrigeration and air-conditioning equipment. The pipeline flags failures 3–4 days ahead and is used every day. The features come from refrigeration-cycle physics (superheat, subcooling, pressures, compressor duty cycle). I separate defrost and door openings from real faults so false alarms stay low. A local LLM explains each forecast in plain language, and colleagues use it through FastAPI and Streamlit tools.
Fukui Megane (EssilorLuxottica) · Fukui
Guffadi Cast · Freelance
District 3292, Zone 11
Personal projects on public data, each with its code on GitHub. Every card says what it does, the result and the tools; “Case study” has the full story.
Finds faults in a CO2 supermarket refrigeration system from its sensor data and names the fault, on open lab data from Oak Ridge National Laboratory. Physics plus ML beat both fixed rules and black-box anomaly detection.
One compressor rack and two display cases (frozen and chilled), with 6 faults put in on purpose. I compare a refrigeration engineer's fault table, black-box anomaly detection on about 150 raw sensors, and a hybrid: 19 physics features from the refrigeration cycle, a small model of what each should be under today's weather, and the gaps read through the same fault table. Scored per fault event, with whole days held out. Found that the black box learned the weather, not the faults, that defrost masking and sensor checks cut false alarms, and that 100% row-level accuracy was only recognising the day. Two faults stayed undetectable, the data has no leak test, and there are only 5 normal days, so I report counts. Personal, 2026.
Predicts how many flights a jet engine has left before it fails, by combining a physics model with machine learning, and still works when almost no failure data exists.
A physics model of a two-spool turbofan (Brayton-cycle gas path), calibrated to NASA's C-MAPSS run-to-failure fleet and used three ways: a digital twin whose grid of Kalman filters tracks compressor damage and predicts remaining life with uncertainty, a simulator that makes failure data with known truth, and physics features for gradient-boosted trees. The hybrid is most accurate; with only 5 failed engines, physics beats ML. Runs on a laptop CPU. Personal, 2026.
Spots defective parts and failing bearings after learning only from normal operation, then runs fast enough on a single factory CPU.
Visual inspection (PatchCore on VisA: PCBs, capsules, cashews) and bearing-vibration monitoring (CWRU), both trained on healthy data only, then shrunk to int8 ONNX and timed on one CPU thread. Found that nearest-neighbour search, not the network, sets edge latency, and that a perfect AUROC still gave 100% false alarms at an unseen motor load until a 2-second recalibration. Personal, 2026.
A Japanese/English search engine for an online store, measured on 1,000 real queries per language and answering in about 32 ms on a laptop CPU.
Japanese and English product search on Amazon's ESCI data: BM25 with MeCab, multilingual E5 embeddings, a tuned hybrid and cross-encoder reranking, both models fine-tuned on ESCI, each scored on 1,000 held-out queries per language. Also tested an open LLM as a relevance judge (it did not match human labels). Plus time-split purchase recommendation (EASE, ALS) and a FastAPI service timed on a laptop CPU. Found that Japanese word segmentation beats any model change, and that "buy it again" beats every learned recommender on repeat purchases. Personal, 2026.
Do classic psychology models or a neural network better predict how people make risky choices? On people it had never seen, the neural network won.
Computational psychiatry on 12,883 Iowa Gambling Task choices from heroin users, amphetamine users and controls (Ahn et al., 2014). Fitted three reinforcement-learning models (PVL-Delta, VPP, ORL) to every person in PyTorch, then ran a fair forecast test: a GRU trained only on other people predicted each person's next 40 choices better than every psychology model. Tested whether decisions reveal addiction (weak signal: heroin partly, amphetamine not), and built an LLM agent that plays the same task with and without chain-of-thought. Personal, 2026.
Re-checked a well-known psychiatry finding with my own code on 1,961 people. It holds up, and adding reaction times cuts the people a future study needs by more than half.
Independent replication of Gillan et al. (2016): do more compulsive people plan less in the two-step task? Re-analysed 1,961 people with my own code and a plan written first, wrote a glmer-style mixed model from scratch, replicated it in a new sample, showed that alcohol problems are not a separate signal, and added a reaction-time marker that halves the sample needed. A full choice + RT drift-diffusion model did worse than the simple version. Personal, 2026.
Finds the exact reply that helped an upset person feel better, and turns many conversations into a searchable catalog of what works. Free to run, no API keys.
Inspired by Andrej Karpathy's HackerTube idea: a model used as an offline data compiler, not a live service. A free, open-source sentiment model (RoBERTa) reads each conversation and finds the first reply after which a distressed person's mood clearly rises; a transparent phrase lexicon names the technique (validation, reflective summarizing, radical acceptance…) and flags dismissive replies like "stay positive" that came first. A Next.js static export serves the catalog on GitHub Pages, and one GitHub Actions workflow processes, commits, builds and deploys. No API keys, no paid services. Sample conversations are synthetic. Personal, 2026.
Detects stress from chest-worn sensor signals, tested only on people the model never saw, the way a real product would be used.
CNN-LSTM and a hybrid deep + physiological-feature model on chest ECG, EDA and breathing signals (WESAD), evaluated only on people the model never saw. Rescued a first version that scored below chance, with an ablation study, seed ensembles and CI-tested code. Personal, 2026.
Sorts microscope images of blood cells into eight types, with an honest look at where it fails and why the score may be optimistic.
Fine-tuned a pretrained Vision Transformer (ViT-Base) on 46,232 microscope images of eight blood-cell types with the Hugging Face Trainer, on a free Colab GPU. Looked at where the 73 test errors fall, not just the headline score, and said why that score is probably optimistic: one merged source ships pre-augmented copies, so near-identical cells can land in both training and test. Personal, 2026.
When a deployed model gets worse, finds which input change caused it: drift tests list the suspects, then each change is undone in turn to see how much of the lost accuracy comes back.
A benchmark of 300 generated card-fraud incidents with known causes: unit mix-ups, missing-value spikes, new customer segments, a rising fraud rate, and traps where a big harmless drift hides a smaller real fault. Ranking columns by drift blamed the harmless column in 60% of traps. The check never named a wrong column; its 8 misses were all "no cause found". A simple drift-times-importance score was just as accurate (99%), so what the check adds is a tested yes or no for each suspect, not higher accuracy. Also run on real electricity-market data (Elec2). No LLM; runs offline on a laptop. Personal, 2026.
Simulated the 2026 World Cup 100,000 times to give every team a title chance. The model's favourite, Spain, won.
Elo ratings feed a win/draw/loss match model with home advantage for the hosts; the whole 48-team tournament (12 groups, best third-placed teams, knockout) is played 100,000 times, and every title chance carries a 95% confidence interval. The model's pre-tournament favourite, Spain, won, and the call I logged before the final scores a Brier of 0.178. A tested Python package with 18 pytest tests and CI. Personal, 2026.
The cricket sibling of the World Cup forecast: two models, one from team results and one ball by ball, play out the rest of the T20 season with the real net-run-rate tie-break and playoff bracket.
Raspberry Pi + OpenCV shelf monitoring with LINE Notify alerts. Led directly to my current role. Fukushima Galilei, 2023–24.
Two transformer-based CV projects: DETR hardhat detection with bipartite matching (loss 1.35), and ViT + knowledge distillation on Oxford Pets (11.6% → 37.6% accuracy). Personal, 2026.
B.E. Electrical & Mechanical Systems
I write up my research as case studies (causal ML, psychometrics, wearables) alongside essays on learning and language — read them here ↗
Open to work · Open to remote collaboration