> SOLUTIONS
Semi-automatic Smoke Segmentation Utility: Initially, a foundation CV model (SAM) transforms the bounding boxes of an existing labelled dataset to segmentation masks in bulk, then the human may approve/denie the masks one-by-one, with convenience through my implemented GUI. |
MedDiSC: Browser-based DICOM de‑identification & Segmentation Tool [1]; implemented for clinical use: de‑identifies and segments bulk DICOM images, curates mask set exported in a single NifTI encoding. |
Startup Initiative: Fire detection system trained on 650 fire images; presented in two competitions and a fire department. |
Jade Maze Solver: JADE-based (Java) agent-oriented system for maze solving; after generating the maze via recursive backtracking. |
DICOM De-identifier Utility: Streamlined a rotation invariant burned-in text removal workflow designed and implemented to de-identify over two million X-Ray DICOM images for INCISIVE (HORIZON 2020) [3]. |
Greek Translator: PyTorch seq2seq language model for Greek–English translation; made up of GRU layers; custom BLEU, trained with teacher forcing scheme; C-accelerated preprocessing (287x); trained on my custom 17k sentences dataset. |
> RESEARCH
Computational Proteomics: Baseline analysis [2] of twelve thousand natural proteomes from UniProtKB; presented in HBioinfo17 (RECOMB 2026). Each of the three visualized columns display a structure aligned set of three proteins sampled from the same cluster. |
CNN Platform: Derived, vectorized, and implemented backpropagation from scratch for any LeNet-like CNN architecture (NumPy/CuPy). This is a proof-of-concept project. The subfigure above shows my model of a convolutional layer; below the way through a rank 3 tensor is indexed. |
M.Eng. Thesis: Transfer learning of LeNet-like Convolutional Neural Networks between image and text modalities through TensorFlow, Keras and NumPy. |
Contrastive Learning Trainer & UI: SimCLR implementation with custom training loop, validation monitoring, checkpointing, and inference. Supports two backbones (LeNet‑style and ResNet) on CIFAR‑10. Achieved improved performance on ResNet compared to LeNet in double the training time. |
Multi-modal Video Summarization: Proof-of-concept pipeline combining audio (MFCC) and visual (CNN) features; keyframe selection formulated as a Knapsack optimization problem; presented as a poster. |
Rigorous Mathematics: The figure shows the Dominated Convergence Theorem for the Lebesgue integral; constructed in alignment with Axler's: Measure, Integration & Real Analysis (2022) book [4]. Click on the figure to view my proof. The purpose of notes such as this is to rigorously fill all the gaps I encounter in mathematical textbooks (somebody has to do it). |
-- provided by --
I am a ML/AI engineer and researcher with a background in deep learning, computational genomics, software engineering, and rigorous mathematics.
My research is comprehensive by default. I build things that work -- in bioinformatics, in engineering, in research, and in the wild. I build clean, scalable, and documented systems, and I'm just as comfortable deriving math as I am shipping code. From backpropagation by hand to pipelines that process 80 million protein sequences, I bridge theory and practice.
Open to freelancing, consulting, and full‑time roles. I deliver value, not just code. I embrace completeness and I refuse to vibe code.
Contact Channels:
Replies within ~1-2 Days.
Expertise:
Familiarity:
Machine Learning Tools: Segment Anything Model, darknet, PyTorch, TensorFlow, keras-ocr
Bioinformatics Tools: DIAMOND, MCL, HipMCL, ChimeraX, BLAST
Fields: Machine Learning, Deep Learning, Linear Algebra, Statistics, Convex Optimization, Mathematical Analysis, Genomics, Finance, Operations Research, Computer Vision, Sequence Processing
- Computational Genomics & Proteomics
- Scalable Graph Clustering & High Performance Computing
- Machine Learning & Mathematical Modeling and Optimization
[1] A. Dimoulakis, E. Politis, M. Nikolaidis, P. Bizopoulos, and K. Votis. "MedDiSC: Medical De-identification, Segmentation & Curation Tool". Manuscript. SSRN. 2024-05-01. Available at: https://ssrn.com/abstract=4813479
[2] A. Dimoulakis. "Computational Genomics -- Analysis of a Scalable and Reproducible Clustering Workflow in Protein Sequence Similarity Networks". Master's Thesis. Aristotle University of Thessaloniki, Greece. 2026. Available at: https://drive.google.com/file/d/1AVaMdupbLhaNpim_sLAkbQsNUr2tf3QD/view?usp=sharing
[3] INCISIVE Project Consortium. "D6.7 Best Practices for DICOM De-identification". Project Deliverable. INCISIVE (HORIZON 2020). 2024-06-30. Available at: https://incisive-project.eu/wp-content/uploads/2024/06/INCISIVE_D6-7_Best-Practices_final.pdf
[4] A. Dimoulakis. "Complementary notes aligned with Axler's Measure, Integration & Real Analysis book, covering foundational topics in metric spaces, Lebesgue integration, and functional analysis". 2022-06. Available at: https://drive.google.com/drive/folders/1p6rbD9gTxQGJupflapUb5uy8YYFP4eRU?usp=sharing
[5] A. Tzanetos, C. Kyriklidis, A. Papamichail, A. Dimoulakis, and G. Dounias. "A Nature Inspired Metaheuristic for Optimal Leveling of Resources in Project Management". Conference Paper. SETN '18: Proceedings of the 10th Hellenic Conference on Artificial Intelligence, 2018-07-09. doi:10.1145/3200947.3201014. Available at: https://doi.org/10.1145/3200947.3201014
"If it isn't documented, it doesn't exist."