Adi Mendelowitz
Senior Data Scientist · ML Researcher
Data scientist, six years in production machine learning across medical AI and energy systems. At GE Healthcare I directed the design of clinical speech and text pipelines, automating how thousands of recorded conversations got tagged by intent. At SolarEdge I ran fleet data for millions of deployed devices: forecasting, survival modelling, and tracing failures back to root cause through Z-score anomaly detection and Pearson correlation.
The work has always pulled toward the research end. These days that means building deep learning architectures from scratch: ResNet and ViT for vision, Mask R-CNN and YOLOv8 for detection and segmentation, SimCLR for self-supervision, and the transformer forecasters PatchTST, iTransformer and TimeMixer, each benchmarked against published results. Most recently: long-horizon forecasting on ETTh1, lesion segmentation on ISIC 2018 medical imaging, and knowledge distillation cutting parameters 65.6x with an 11x inference speedup. The forecasting thread became a sole-author paper on channel dependence in PatchTST, accepted at Transactions on Machine Learning Research (2026).
B.Sc. Mathematics. The code is on GitHub.
Independent Research
Computer Vision, Time Series & Deep Learning
Publication · Transactions on Machine Learning Research, 2026 (accepted)
Equal Accuracy, Unequal Cost: Channel Dependence in PatchTST under Controlled Coupling
Sole author · PyTorch · Kaggle T4 · time-series forecasting · Transformers · experimental design · reproducibility
Channel-independent and channel-dependent PatchTST are usually compared on observational benchmarks where correlation strength, dimensionality and temporal dynamics vary together. This paper isolates them one at a time with controlled synthetic generative processes (a factorial AR(1) grid, a leader-follower VAR(1) family with tunable lag-1 coupling, and a block-covariance family), with ETTh1 and ECL as real-data context. Across every tested cell the CD-CI accuracy difference stays inside a pre-specified 1% band, and a closed-form analysis of the generative model shows why: at the evaluation horizon no predictor, CI or CD, has any cross-channel signal left to exploit. Cost is the real difference. CD needs smaller batches and far more gradient steps, and at 321 variates a CD epoch costs 6.7 times a CI epoch. CI is the better default on accuracy per unit compute. The repository ships a one-command reproduction: it checks every result file against a committed checksum, reruns the analysis, and regenerates every table and figure in under a minute on a CPU. Every version of the code, including the one the paper cites, is archived on Zenodo under a DOI.
CD-CI within 1% MSE, every cell
CD epoch 6.7x CI at 321 variates
639 training runs, 5 seeds per cell
Closed-form ceilings match to float64
Time series forecasting with transformers
PyTorch · ETTh1 · Kaggle T4
PatchTST, iTransformer, and TimeMixer built from scratch and evaluated on ETTh1 across four horizons against a linear baseline, with a unit-test suite covering every model.
3 architectures from scratch
PatchTST MSE 0.4599 at H=96 (5 seeds)
4 horizons
Architecture implementations
PyTorch · torchvision · timm · CIFAR-10 · MNIST
ResNet-18 and ViT-Tiny built from scratch, including multi-head self-attention, plus a NumPy CNN with hand-derived backpropagation. SimCLR contrastive pretraining reproduced the original paper's label-efficiency results.
ResNet-18 93.4%
ViT-Tiny 86.7%
SimCLR 68.2% linear eval
61.6% at 1% labels
CNN in NumPy 90.9% MNIST
Adversarial robustness
PyTorch · CIFAR-10 · ResNet-18
FGSM, PGD, and Carlini-Wagner attacks implemented from scratch against a PGD adversarial-training defence, verified against gradient masking via all four Athalye et al. diagnostic checks.
78.0% clean / 47.2% robust accuracy
C&W distortion 0.231 → 0.717 pre/post defence
4/4 Athalye checks passed
Medical image segmentation
Mask R-CNN · ISIC 2018 · Kaggle T4
Mask R-CNN fine-tuned for skin lesion segmentation on ISIC 2018 Task 1.
Jaccard 0.782
Object detection
YOLOv8 · PCB defect dataset · Kaggle T4
YOLOv8n fine-tuned on a public PCB defect dataset with domain-specific augmentation.
mAP@0.5 0.9896
Model compression
PyTorch · INT8 quantization · Knowledge distillation
Static INT8 quantization and knowledge distillation benchmarked against an FP32 baseline for latency, throughput, and accuracy.
3.95x size reduction
1.52x inference speedup
65x compression
Professional Experience
SolarEdge Technologies
Mar 2022 – Aug 2025
Senior Data Science Engineer · Fleet Data Team Lead · Data Scientist
- Built multilingual translation pipelines for thousands of recorded customer-support conversations, using TF-IDF term-relevance scoring and hierarchical clustering to triage which terms needed translation and cut LLM call volume; validated output quality with domain experts.
- Led failure and survival analysis across device fleets, including early-life (infant-mortality) and survival-rate reliability tracking; improved preemptive failure detection by 25% and reduced maintenance costs by 20%.
- Represented fleet data on a thrice-weekly cross-departmental working group resolving the most severe inverter malfunctions across 3 models with millions of deployed units each; drove root cause analysis, severity assessment, and staged rollout testing from thousands to hundreds of thousands of live devices before fleet-wide rollout.
- Built the telemetry pipelines monitoring millions of inverters and battery units at 24-hour latency, feeding Z-score-based anomaly detection and Pearson-correlation root cause attribution run over noisy, correlated multivariate sensor signals.
- Flagged critical software bugs from log and telemetry analysis, with 98% of flagged issues confirmed as real defects.
- Built a standardised benchmark database covering product performance and end-of-life metrics, streamlining cross-team analysis, reliability assessments, and maintenance planning.
- Directed delivery of a Python/Streamlit SQL interface, adopted across hardware, inverter, and finance teams so non-analyst managers could query fleet data without analyst support.
- Built a real-time thermal estimation system for inverter devices, cutting overheating incidents by 30%; reduced product failure rates by 60% through systematic root cause analysis.
- Delivered simulation-based competitor analysis informing $10M+ in strategic investment decisions.
- Co-founded the SE Data Forum, building cross-company collaboration among 30+ data professionals.
- Mentored a cross-functional team of 4-5; established Agile practices and code quality standards.
GE Healthcare – Life Care Solutions
Nov 2019 – Jan 2022
Data Operations Engineer
- Directed the design of a multi-step pipeline for clinical speech and text data across hospital sites, and the automation of a proprietary tagging tool by intent domain, creating a more intuitive and cost-effective process.
- Managed the continuous learning cycle of ML/DL models using performance diagnostics against a 10,000+ utterance ground truth set, as well as reviews by the tagging team.
- Led a team of 2–4 developers and data science interns, plus 10+ remote clinical domain experts.
Zeek Mobile – Marketplace for Gift Cards
May 2018 – Jan 2019
Junior Data Scientist
- Implemented feature selection methods in Python and Node.js to address overfitting.
- Ran SQL-based exploratory analysis and presented findings via Periscope.
Education
B.Sc. Mathematics
The Open University of Israel
Part-time while employed full-time.
Contact
English C1 (IELTS 8.0)