Alexandru Mihai

Research Data Scientist

Alexandru Mihai

Bridging applied mathematics, computer vision, and machine learning

  • Ph.D. Applied Mathematics
  • 8+ years in research
  • Okinawa, Japan
View projects Download CV
EXPLORE
About

Professional Summary

Research Data Scientist with a Ph.D. in Applied Mathematics and 8+ years building computational solutions across diverse scientific domains at an international research institute in Japan. I am the sole data scientist for a marine climate-change research unit, independently designing and delivering ML pipelines, satellite-imagery workflows, computer-vision systems, NLP tools, and data infrastructure that support active research and publication.

My work spans the full project lifecycle — from stakeholder consultation and problem definition through implementation, validation, documentation, and end-user training. I am now looking to bring this breadth, rapid domain adaptability, and track record of shipping working solutions to a product-oriented industry team.

Portrait of Alexandru Mihai
Selected work

Featured Projects

Selected works in AI, environmental monitoring, and computer vision — with the metrics that matter.

UMAP embedding of fish images shown as a 2D map of image tiles
Deep Learning

DEEPFIN — Fish Image Classification from Unlabeled Data

An unsupervised deep-learning pipeline that classifies fish images with no labels, combining CNN feature extraction, UMAP dimensionality reduction, and HDBSCAN clustering. Validated on clownfish developmental staging and the Fish4Knowledge benchmark, and released as a pip-installable Python package with a CLI.

0.873 ARIpip installable package
  • PyTorch
  • CNN feature extraction
  • UMAP
  • HDBSCAN
  • Python
  • CLI
GitHub
Orthophoto of an Okinawan mangrove with clustered vegetation classes overlaid
Remote Sensing

Satellite-Based Blue Carbon Monitoring

Satellite-imagery analysis for mangrove ecosystem monitoring in Okinawa — integrating Sentinel-2 optical data, LIDAR point clouds, and drone imagery for change detection, biomass estimation, and carbon-sequestration quantification. Includes an end-to-end Sentinel-2 change-detection pipeline implementing six methods with seam correction, Otsu thresholding, polygon vectorization, and Leaflet visualization.

6 change-detection methodsS-2 + LIDAR data fusion
  • Google Earth Engine
  • Sentinel-2
  • LIDAR
  • GDAL
  • QGIS
  • Leaflet
  • Python
View interactive map Temporal analysis
UMAP 2D embedding of macrophyte image tiles, coloured by predicted species
Computer Vision

LeafMosaic — Aquatic Macrophyte Classification at Scale

A tile-based CNN pipeline classifying four aquatic macrophyte species (Lemna minor, Spirodela polyrhiza, Azolla filiculoides, Salvinia natans) using DenseNet-121, VGG-16, and a custom CNN with transfer learning, across a 69-unit mesocosm experiment. Published as open-source software; associated ecology manuscript in preparation.

91.8% mean val. accuracy19,754 image tiles
  • DenseNet-121
  • VGG-16
  • Transfer learning
  • PyTorch
  • Albumentations
  • Python
GitHub Zenodo (DOI)
Doctoral & ongoing research

Research Showcase

Soap Film Mediated 3D Self-Assembly

Suspended and Displacement-Driven Geometries

  • OIST
  • 2017 – 2022
  • Supervisor: Prof. Eliot Fried

My doctoral research explored how soap films can act as a medium for the self-assembly of complex 3D structures. By suspending specially designed centimetre-scale tiles in soap films and carefully controlling the experimental conditions, we observed the spontaneous formation of platonic solids and other geometric structures. The work combined experimental physics with custom image analysis and mathematical modelling to understand the principles governing these self-assembly processes.

  • First experimental demonstration of soap-film-mediated 3D self-assembly
  • Developed a mathematical framework for predicting assembly outcomes
  • Built bespoke experimental rigs with automated, high-throughput image capture
  • Wrote image-processing and analysis algorithms from scratch in Python / C++
  • Applied Mathematics
  • Experimental Physics
  • Image Analysis
  • Optimization
  • Python
  • C++
Read the thesis (PDF)

Current research at OIST

NLP & Biocultural Analysis

Applied LLM-based text embeddings and sentiment analysis to anthropological interview data from Okinawan fishing communities (Ikei Jima, Itoman, Onna Village), extracting biocultural indicators of marine-ecosystem health and building semantic-network visualizations of community-level differences.

  • Sentence embeddings
  • Sentiment analysis
  • TF-IDF
  • Semantic networks

LLM Uncertainty Research

Developed training-free confidence-scoring systems that read hidden-layer activations across multiple transformer families (Mistral-7B, Llama variants, Phi-3 Mini), evaluated across temperature settings to estimate when a model is — and is not — reliable.

  • Mistral-7B
  • Llama
  • Phi-3 Mini
  • Uncertainty quantification

Coral Bleaching Analysis

Built interactive Python and JavaScript visualizations for K-means cluster colour data and CoralWatch bleaching scores across control / treatment mesocosm groups, including a custom bleaching index and multi-panel comparative figures.

  • K-means
  • Custom bleaching index
  • Plotly
  • JavaScript
Open the dashboard

Microscopy & Genomics

3D cell segmentation of z-stack microscopy images with Cellpose, plus bioinformatics pipelines for fish-genome methylation studies and an evaluation of HPC-versus-workstation requirements for genomics workflows.

  • Cellpose
  • 3D segmentation
  • Methylation analysis
  • SLURM/HPC

3D Printing & Fabrication

Design and fabrication of custom research equipment — including passive eDNA samplers for marine field deployment and reusable wet-lab tools — reducing costs and enabling rapid prototyping for the unit.

  • CAD
  • 3D printing
  • Rapid prototyping
  • Field hardware

Data Visualization & Infrastructure

Developed and maintained interactive web-based visualization platforms for eDNA and ecological data, and manage the unit’s data infrastructure: on-site backup servers, 100TB+ flash/bucket storage, and SLURM-scheduled HPC.

  • Leaflet
  • PostGIS
  • SpatiaLite
  • 100TB+ storage
  • SLURM
Mangrove report
For the love of it

Programming Projects

Algorithms, simulations, and reinforcement learning, built outside of funded research.

Learned weight matrix revealing corner-favouring spatial preferences

Optimal Strategy for 2048 via Reinforcement Learning

Q-learning + computer vision

A Q-learning agent that discovers a non-obvious corner-favouring strategy for 2048, with the board state read directly from the screen by a small computer-vision pipeline. The learned weight matrix reveals emergent spatial preference patterns.

  • Reinforcement Learning
  • Q-learning
  • Computer Vision
  • Python
Read the write-up (PDF)
Space-time diagram showing emergent traffic jams at high dawdle probability

Nagel–Schreckenberg Traffic on Road Networks

Cellular automata

A cellular-automaton traffic model running on arbitrary road-network topologies. Space–time diagrams reveal how emergent traffic jams form spontaneously as the random-braking ("dawdle") probability rises.

  • Cellular Automata
  • Network Analysis
  • Simulation
  • Python
Streamline visualization of a Lattice-Boltzmann flow simulation

Lattice-Boltzmann CFD Simulations

Computational fluid dynamics

Fluid-flow simulations using the Lattice-Boltzmann method, including airfoil and channel flows across Reynolds numbers, with streamline visualization and integration with the OpenLB framework for 3D simulations.

  • Lattice Boltzmann
  • OpenLB
  • Numerical Methods
  • C++
  • Parallel Computing

Blood Flow in Stenotic Vessels

Biomedical modelling

Real-time 2D/3D blood-flow simulations characterizing flow patterns in stenotic (narrowed) vessel geometries — work begun during a research assistantship at Fraunhofer MEVIS toward CFD-assisted medical diagnosis.

  • CFD
  • Biomedical Imaging
  • Numerical Methods
  • Scientific Computing
Output

Publications & Software

Peer-reviewed work and open-source tools.

Papers

  1. In review iScience (Cell Press)

    DEEPFIN: A Deep Learning Tool for Fish Image Classification from Unlabeled Data

    Mihai, A., et al.

    Unsupervised classification of fish images via CNN feature extraction, UMAP, and HDBSCAN — validated on clownfish developmental staging and the Fish4Knowledge benchmark (ARI 0.873).

    GitHub
  2. In preparation

    Aquatic Macrophyte Mesocosm Classification with the LeafMosaic Pipeline

    Mihai, A., et al.

    Tile-based CNN classification of four aquatic macrophyte species across a 69-unit mesocosm experiment (91.8% mean validation accuracy over 19,754 tiles).

Software

  1. Released Python package

    DEEPFIN

    Open-source, pip-installable Python package with a CLI for unsupervised fish-image classification.

    GitHub
  2. Released Research software

    LeafMosaic

    Automated image-classification pipeline for aquatic plant species, published with a citable DOI on Zenodo.

    GitHub Zenodo · DOI 10.5281/zenodo.20421921
Career

Experience

  1. Senior Research Data Scientist

    2022 – Present

    Okinawa Institute of Science and Technology (OIST) — Marine Climate Change Unit

    Okinawa, Japan

    Sole data scientist for a marine climate-change research unit, responsible for all computational analysis, ML development, data infrastructure, and technical support across a portfolio of concurrent projects — spanning data science, software engineering, geospatial analysis, hardware fabrication, and fieldwork support.

    • Built DEEPFIN (CNN + UMAP + HDBSCAN) and LeafMosaic (DenseNet/VGG), released as open-source packages
    • Lead satellite remote-sensing for blue-carbon monitoring (Google Earth Engine, Sentinel-2, LIDAR, drone)
    • Applied LLM embeddings, sentiment analysis and uncertainty estimation to research data
    • Manage 100TB+ storage and SLURM/HPC infrastructure; mentor students on ML and remote sensing

  2. Research Scientist / Ph.D. Candidate

    2017 – 2022

    Okinawa Institute of Science and Technology (OIST)

    Okinawa, Japan

    Investigated self-assembly systems within soap films, independently managing the full research pipeline: question discovery, experimental design, prototyping, data collection, imaging, image processing, quantitative analysis, and scientific writing. Developed mathematical and computational models for interfacial fluid dynamics and emergent geometric phenomena.

  3. Teaching Assistant — Computational Mathematics

    2020

    Summer@ICERM, Brown University

    Remote

    Mentored student teams on computational-mathematics topics including graph algorithms, spectral clustering, and iterative optimization methods; reviewed Python code architecture and guided reproducible-research practices.

  4. Research Assistant — Medical Image Processing

    2014 – 2015

    Fraunhofer Institute for Digital Medicine MEVIS

    Bremen, Germany

    Developed image-processing and analysis algorithms for biomedical imaging datasets in a clinical research environment, and completed a bachelor’s thesis on blood-flow dynamics simulations in collaboration with MEVIS researchers.

Toolkit

Skills & Expertise

Programming

  • Python
  • C / C++
  • R
  • SQL
  • JavaScript / TS
  • MATLAB

Machine Learning & Deep Learning

  • PyTorch
  • TensorFlow
  • CNNs (DenseNet, VGG, ResNet)
  • Transfer learning
  • UMAP / HDBSCAN / k-means
  • Model calibration & validation
  • Uncertainty quantification

Computer Vision

  • 2D / 3D segmentation (Cellpose)
  • Object classification
  • Image tiling
  • Feature extraction
  • Morphological analysis
  • Automated CV pipelines

NLP & LLMs

  • LLM text embeddings
  • Sentiment analysis
  • TF-IDF
  • Semantic network analysis
  • LLM uncertainty estimation

Satellite & Geospatial

  • Google Earth Engine
  • Sentinel-2
  • LIDAR point clouds
  • Change detection (MAD, PCA, Otsu)
  • QGIS / GDAL
  • PostGIS / SpatiaLite / Leaflet

Data & Infrastructure

  • PostgreSQL
  • ETL & data validation
  • Large-scale datasets (100TB+)
  • Docker / Linux
  • SLURM / HPC
  • Git / GitHub
Off the clock

Photography

A long-running habit of looking closely — film, water, waves, and skies.

35mm Film

Paris

Water Drops

Waves

Clouds