ML engineer with a PhD. Research, side projects, and open source on self-supervised learning, LLMs, and computer vision. (Feed contents: blog, changelog, publications, projects.)https://gkos.dev/https://arxiv.org/abs/2606.31201https://arxiv.org/abs/2606.31201ExPLoRe repurposes Soft Mixture of Experts (MoE) dispatch weights as learned, per-patch loss coefficients for multi-objective masked image modeling. Loss-coupling lets gradients flow through dispatch weights to the router, enabling content-dependent specialization where different patches receive different emphases across training objectives. ViT-Base + CLIP-B/16 teacher reaches 80.6% linear probe and 85.3% finetuning accuracy on ImageNet-1K.Thu, 31 Dec 2026 23:59:59 GMTECCV 2026https://arxiv.org/abs/2603.29009https://arxiv.org/abs/2603.29009MEDiC introduces a multi-objective framework for distilling knowledge from CLIP into smaller, task-specific models for remote sensing applications. By jointly optimizing multiple objectives during distillation, the method achieves strong performance on downstream tasks while significantly reducing computational requirements.Thu, 31 Dec 2026 23:59:59 GMTarXiv preprinthttps://gkos.dev/blog/per-patch-loss-coupling/https://gkos.dev/blog/per-patch-loss-coupling/A walkthrough of ExPLoRe, our ECCV 2026 paper on per-patch loss weighting for masked image modeling. It covers the idea, the main ablations, and the parts I'm still unsure about.Wed, 24 Jun 2026 12:00:00 GMTresearchmlhttps://gkos.dev/changelog/#2026-06-17-stats-rewired/https://gkos.dev/changelog/#2026-06-17-stats-rewired/Swapped the analytics stack from Cloudflare beacons + Umami Cloud to PostHog (free tier, EU region). Cloudflare was getting ad-blocked, Umami's API moved behind a paywall. The Traffic section reads real numbers again. Bonus: a new CV downloads counter, and a site-wide accessibility and mobile-polish pass.Wed, 17 Jun 2026 08:00:00 GMThttps://gkos.dev/changelog/#2026-06-16-phd-blog-post/https://gkos.dev/changelog/#2026-06-16-phd-blog-post/Six lessons from five and a half years of an ML PhD. None of them about the technical content: that part is in the papers. This is the rest.Tue, 16 Jun 2026 15:00:00 GMThttps://gkos.dev/blog/what-the-phd-gave-me/https://gkos.dev/blog/what-the-phd-gave-me/Five and a half years of an ML PhD, six lessons that aren't in any of the papers. The kind of things I keep telling friends still in the program.Mon, 15 Jun 2026 12:00:00 GMTcareerpersonalresearchhttps://gkos.dev/changelog/#2026-06-05-take-them-back/https://gkos.dev/changelog/#2026-06-05-take-them-back/Hit a reaction by accident? Tap it again, it's gone. Same for comments: if you posted one from this device, you'll see a Delete button next to it. No login needed: tracked per device, not per account.Fri, 05 Jun 2026 13:00:00 GMThttps://gkos.dev/changelog/#2026-06-04-blog-comments/https://gkos.dev/changelog/#2026-06-04-blog-comments/Every blog post now has a comments thread at the bottom. New comments feed into the stats page just like reactions, with a per-country breakdown.Thu, 04 Jun 2026 21:00:00 GMThttps://gkos.dev/changelog/#2026-06-04-smaller-lighter-site/https://gkos.dev/changelog/#2026-06-04-smaller-lighter-site/Cleaned up a few half-built things: removed the experimental Hero Lab and the internal widgets showcase, renamed the level editor to /explore-editor, and tightened a couple of stale routes.Thu, 04 Jun 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-06-03-stats-properly-wired/https://gkos.dev/changelog/#2026-06-03-stats-properly-wired/Every card on the stats page now pulls real data: site views from Cloudflare, top pages from Umami, reactions from the new backend, contributions from GitHub, downloads from PyPI and HuggingFace. No more placeholders.Wed, 03 Jun 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-06-02-blog-reactions/https://gkos.dev/changelog/#2026-06-02-blog-reactions/Each blog post now has an emoji picker at the end. 👍 ❤️ 🎉 💡 — your reactions accumulate and show up on the stats page.Tue, 02 Jun 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-05-22-home-less-duplication/https://gkos.dev/changelog/#2026-05-22-home-less-duplication/Resolved the two "Workbench" widgets on the home page. Now there's one Workbench card (the tools I use) and one Projects card (the work I've built). Each links to its own page.Fri, 22 May 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-05-15-projects-tightened/https://gkos.dev/changelog/#2026-05-15-projects-tightened/The projects page leads with the strongest work now. Reordered featured, renamed "ML Research" to "Machine Learning," and moved older educational repos (the Numpy CNN, the p5 shooter) to the bottom.Fri, 15 May 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-05-03-inspirations-curated/https://gkos.dev/changelog/#2026-05-03-inspirations-curated/The inspirations page is now a real list of who and what I actually return to: mentors, researchers, books on the shelf, YouTube channels, podcasts, newsletters, and blogs. About 34 items, each with a link.Sun, 03 May 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-04-28-about-redesign/https://gkos.dev/changelog/#2026-04-28-about-redesign/The about page has a new shape: work history first with a rail timeline, then the story behind it in 8 chapters. A scroll-tracked curve weaves through every chapter.Tue, 28 Apr 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-04-26-blog-projects-publications/https://gkos.dev/changelog/#2026-04-26-blog-projects-publications/Redesigned the Blog, Projects, and Publications pages. Compact lists, click-to-expand details, and proper co-author and citation data pulled live.Sun, 26 Apr 2026 12:00:00 GMThttps://gkos.dev/changelog/#2026-04-25-contact-wall-changelog/https://gkos.dev/changelog/#2026-04-25-contact-wall-changelog/Rebuilt the Contact, Community Wall, and Changelog pages from the ground up. Cleaner layouts, better typography, and the new widgets in their natural habitat.Sat, 25 Apr 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-04-24-widgets-redesign/https://gkos.dev/changelog/#2026-04-24-widgets-redesign/Reworked every widget on the site for a consistent look and feel: same chrome, same spacing, same hover states. A few new ones too.Fri, 24 Apr 2026 10:00:00 GMThttps://gkos.dev/changelog/#2026-04-23-explore-soon/https://gkos.dev/changelog/#2026-04-23-explore-soon/The interactive version of the portfolio is nearly ready. Teaser hints live in the Hello World post.Thu, 23 Apr 2026 09:00:00 GMThttps://gkos.dev/changelog/#2026-04-22-math-in-blog/https://gkos.dev/changelog/#2026-04-22-math-in-blog/Blog posts can now render LaTeX equations through KaTeX. Inline and block math both work.Wed, 22 Apr 2026 12:00:00 GMThttps://gkos.dev/changelog/#2026-04-22-first-blog-posts/https://gkos.dev/changelog/#2026-04-22-first-blog-posts/Shipped Hello World and How I got here.Wed, 22 Apr 2026 11:00:00 GMThttps://gkos.dev/changelog/#2026-04-22-site-launch/https://gkos.dev/changelog/#2026-04-22-site-launch/The new portfolio is live. The first post, Hello World, walks through the history and what's new.Wed, 22 Apr 2026 10:00:00 GMThttps://gkos.dev/blog/how-i-got-here/https://gkos.dev/blog/how-i-got-here/The long version of my background, from Greece to a PhD in Tennessee, with a few detours in between.Mon, 20 Apr 2026 04:00:00 GMTpersonalhttps://gkos.dev/blog/hello-world/https://gkos.dev/blog/hello-world/A fresh site, a fond farewell to the old one, and a small tease of what's coming next.Wed, 15 Apr 2026 12:00:00 GMTpersonalmetahttps://github.com/aicip/ExPLoRehttps://github.com/aicip/ExPLoReOfficial PyTorch implementation of the ECCV 2026 ExPLoRe paper. Repurposes Soft-MoE dispatch weights as learned, per-patch loss coefficients for multi-objective masked image modeling — loss-coupling lets gradients flow through dispatch to the router for content-dependent specialization. ViT-Base + CLIP-B/16 teacher hits 80.6% linear probe and 85.3% finetune on ImageNet-1K.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPyTorchSelf-SupervisedSoft-MoEMIMCLIPComputer-VisionECCV-2026https://huggingface.co/drkostas/MEDiC-ViT-Basehttps://huggingface.co/drkostas/MEDiC-ViT-BaseOfficial PyTorch implementation of MEDiC: Multi-objective Exploration of Distillation from CLIP. Combines token distillation, CLS alignment, and pixel reconstruction with Evolved Part Masking. Achieves 85.07% finetuning and 73.92% k-NN on ImageNet-1K.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPyTorchSelf-SupervisedCLIPMIMComputer-VisionHuggingFacehttps://github.com/aicip/Cross-Scale-MAEhttps://github.com/aicip/Cross-Scale-MAEOfficial code for the paper 'Cross-Scale MAE: A Tale of Multi-Scale Exploitation in Remote Sensing'. Self-supervised learning for multi-scale geospatial imagery.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPyTorchMIMComputer-Visionhttps://huggingface.co/drkostas/MaskDistill-ViT-Basehttps://huggingface.co/drkostas/MaskDistill-ViT-BaseFirst open PyTorch reproduction of MaskDistill with pre-trained weights. Reproduces 84.8% finetuning accuracy (paper: 85.3%), with full evaluation suite: semantic segmentation, object detection, k-NN, and linear probe.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPyTorchSelf-SupervisedCLIPComputer-VisionHuggingFacehttps://github.com/drkostas/Minecraft-AIhttps://github.com/drkostas/Minecraft-AIA Reinforcement Learning agent that learns how to solve maze missions in Minecraft.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPyTorchReinforcement-LearningMinecrafthttps://github.com/drkostas/3D-Semantic-Segmentationhttps://github.com/drkostas/3D-Semantic-SegmentationSemantic Segmentation with Transformers on 3D Medical Images.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPyTorchOpenCVSegFormerSemantic-SegmentationMedical-Imaginghttps://github.com/drkostas/Bert-Question-Answeringhttps://github.com/drkostas/Bert-Question-AnsweringBERT-based question answering / reading comprehension methods on Rinehart Novels.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPyTorchSpaCyHuggingFaceTransformersBERThttps://github.com/drkostas/accident-severity-predictionhttps://github.com/drkostas/accident-severity-predictionPredicting the severity of car accidents from various attributes.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningPandasScipyBayesian-OptimizationXGBoostNeural-Networkhttps://github.com/drkostas/covid19-vaccinations-predicthttps://github.com/drkostas/covid19-vaccinations-predictSimultaneous Time Series Forecasting on the global COVID-19 Daily Vaccinations.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningTensorflowLSTMsMultivariate-Time-Serieshttps://github.com/drkostas/Insta-Likes-Predicthttps://github.com/drkostas/Insta-Likes-PredictFirst attempt on predicting the likes a photo will get on Instagram.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningTensorflowOpenCVInstagramScraperCNNhttps://github.com/drkostas/RL-Value-Iterationhttps://github.com/drkostas/RL-Value-IterationImplementation of value iteration algorithm for calculating an optimal MDP policy.Wed, 15 Apr 2026 12:00:00 GMTMachine LearningMarkov-Decision-ProcessValue-IterationRLhttps://gkos.devhttps://gkos.devThe personal portfolio + blog you're on right now. Astro + Tailwind + React islands + Supabase + PostHog. Built to ship fast, stay light, and own its own analytics.Wed, 15 Apr 2026 12:00:00 GMTProductsAstroTailwindReactSupabasePostHogVercelhttps://hevy2garmin-demo.gkos.devhttps://hevy2garmin-demo.gkos.devWeb tool that pulls strength workouts from Hevy and writes them into Garmin Connect with exercise names, sets, reps, weights, HR overlay, and calorie estimation. Solves a real gap in both apps' APIs. 36 stars, 109 forks — the most-forked project I have.Wed, 15 Apr 2026 12:00:00 GMTProductsPythonFastAPIGarminHevyOAuthhttps://fleetsmart.aihttps://fleetsmart.aiAI-powered maritime decision engine. Helps shipowners make profitable positioning decisions: which vessel goes where, when to reposition, how to bid on cargo. Role: technical lead on a contract build — backend, LLM agent layer, data pipelines, cloud infra. Closed-source (commercial product), live in production with paying customers.Wed, 15 Apr 2026 12:00:00 GMTProductsFastAPINext.jsGCPLLMAgentsPostgreSQLhttps://gkos.dev/projects/#shiftmd/https://gkos.dev/projects/#shiftmd/Intelligent shift scheduling system for medical departments using constraint programming optimization.Wed, 15 Apr 2026 12:00:00 GMTProductsNext.jsPythonOR-ToolsSupabasehttps://xpensai.comhttps://xpensai.comAI-powered expense management platform with automated receipt scanning, OCR, and intelligent categorization.Wed, 15 Apr 2026 12:00:00 GMTProductsPythonAWSAzureGPT-4oServerlesshttps://soma-demo.gkos.devhttps://soma-demo.gkos.devPersonal health and fitness dashboard aggregating data from Garmin, Strava, and Hevy into a unified analytics view.Wed, 15 Apr 2026 12:00:00 GMTProductsPythonNext.jsGarminStravahttps://search.gkos.devhttps://search.gkos.devA search engine for the Gutenberg Project archive, written in Python with a Flask front-end. Inverted-index lookup, ranked retrieval, query suggestions.Wed, 15 Apr 2026 12:00:00 GMTToolsPythonFlaskIRSearchhttps://github.com/drkostas/garmin-authhttps://github.com/drkostas/garmin-authSelf-healing Garmin Connect OAuth authentication. Handles the complex SSO flow (OAuth1 to OAuth2), automatic token refresh, and rate limit recovery.Wed, 15 Apr 2026 12:00:00 GMTToolsPyPiGarminOAuthPythonauthenticationhttps://github.com/drkostas/high-sqlhttps://github.com/drkostas/high-sqlA high-level SQL command utility. Currently only MySQL is supported.Wed, 15 Apr 2026 12:00:00 GMTToolsPyPiMySQLCircleCIwrapperhttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&citation_for_view=b___QQ8AAAAJ:Tyk-4Ss8FVUChttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&citation_for_view=b___QQ8AAAAJ:Tyk-4Ss8FVUCThis work presents a trustworthy AI approach for early dementia detection, focusing on robust feature masking techniques and clinical interpretability to ensure reliable and transparent diagnostic support systems.Wed, 31 Dec 2025 23:59:59 GMTIEEE/ACM CHASE 2025https://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&citation_for_view=b___QQ8AAAAJ:UeHWp8X0CEIChttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&citation_for_view=b___QQ8AAAAJ:UeHWp8X0CEICThis paper introduces a novel approach to semi-supervised object detection by incorporating a teaching assistant mechanism to improve the quality of pseudo-labels, enhancing overall model performance and reliability.Wed, 31 Dec 2025 23:59:59 GMTRevista Tecnología en Marchahttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=3&citation_for_view=b___QQ8AAAAJ:qjMakFHDy7sChttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=3&citation_for_view=b___QQ8AAAAJ:qjMakFHDy7sCWe reformulate construction phase classification as a transition detection problem and introduce a Koopman-Based Transition Detection (KTD) method that applies Koopman operator theory to analyze the nonlinear dynamics of material histograms in a linear framework. KTD employs a sliding window with Dynamic Mode Decomposition (DMD) on time-series material histograms and detects transition points by analyzing eigenvalue movement. Compared to CNN-based methods, KTD demonstrates enhanced accuracy and reduced temporal error.Tue, 31 Dec 2024 23:59:59 GMTIEEE/IGARSS 2024https://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=3&citation_for_view=b___QQ8AAAAJ:2osOgNQ5qMEChttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=3&citation_for_view=b___QQ8AAAAJ:2osOgNQ5qMECThis research focuses on improving the fine-tuning process of self-supervised learning models for remote sensing, particularly the Cross-Scale Masked Auto-Encoder (MAE). We tackle the challenges of intricate, multi-source imagery and present advancements in adapting the Cross-Scale MAE for diverse remote sensing environments.Tue, 31 Dec 2024 23:59:59 GMTIEEE/IGARSS 2024https://arxiv.org/abs/2402.00689https://arxiv.org/abs/2402.00689We conduct a comparative analysis of four advanced LLMs (GPT-3.5, GPT-4, Bard, Gemini) across 9 tasks to assess code generation capabilities. We focus on identifying conditions under which LLMs can be effectively and safely deployed for code generation, with emphasis on security awareness via two distinct developer personas. We collected 61 code outputs and analyzed them across functionality, security, performance, complexity, and reliability.Tue, 31 Dec 2024 23:59:59 GMTarXiv preprinthttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=1&citation_for_view=b___QQ8AAAAJ:d1gkVwhDpl0Chttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=1&citation_for_view=b___QQ8AAAAJ:d1gkVwhDpl0CWe present Cross-Scale MAE, a self-supervised model built on the Masked Auto-Encoder (MAE) framework for remote sensing image understanding. Cross-Scale MAE employs scale augmentation and enforces cross-scale consistency through both contrastive and generative losses to ensure consistent and meaningful representations for downstream tasks. Experimental evaluations demonstrate superior performance compared to standard MAE and other state-of-the-art remote sensing MAE methods.Sun, 31 Dec 2023 23:59:59 GMTNeurIPS 2023https://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=1&citation_for_view=b___QQ8AAAAJ:u-x6o8ySG0sChttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=1&citation_for_view=b___QQ8AAAAJ:u-x6o8ySG0sCWe exploit self-supervised contrastive learning for semantic segmentation in aerial imagery. In addition to feature-level CL, we add another level of contrastive learning at the semantic level, taking advantage of segmentation output. We embed local mutual information in the semantic-level CL to enforce local consistency, enhancing representation power and generalization. The proposed multi-level contrastive learning with local consistency (mCL-LC) shows superior performance and better generalization, especially under domain shift.Sun, 31 Dec 2023 23:59:59 GMTIEEE/WACV 2023https://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=1&citation_for_view=b___QQ8AAAAJ:u5HHmVD_uO8Chttps://scholar.google.com/citations?view_op=view_citation&hl=en&user=b___QQ8AAAAJ&sortby=pubdate&authuser=1&citation_for_view=b___QQ8AAAAJ:u5HHmVD_uO8CWe combine network topology properties (loose similarity and local edge betweenness, an alternative to Girvan-Newman's edge betweenness) with intrinsic user content information to introduce a novel and highly distributed hybrid community detection methodology. The proposed approach is tested on real social graphs and compared to classic divisive community detection algorithms, proving exceptionally scalable, highly efficient, and accurate in revealing the subjacent network hierarchy.Tue, 31 Dec 2019 23:59:59 GMTAlgorithms (MDPI) 2019