I'm Kostas. Here's my story.

Designed and built FleetSmart's AI decision platform for maritime fleet operators: a multi-objective optimization engine that ranks vessel-opportunity matches across cost, speed, risk, and environmental targets, with A/B-testable scoring architectures and a fleet KPI engine for benchmarking and anomaly detection.

Architected the data and reasoning stack: 9+ real-time pipelines (dual-provider AIS vessel tracking, Baltic Exchange bunker prices, FIS Live forward freight curves, MABUX fuel spreads, AXS vessel profiles and port congestion, news sentiment), multi-provider LLM tool calls (Gemini / OpenAI / Anthropic via MCP), voyage simulation for what-if planning, and post-voyage cost feedback loops.

Architected a multi-agent LLM framework for natural language to executable code, achieving 82% functional success on smart home automations and 87.6% on MBPP.

Engineered an impossibility detector that flagged over 93% of unfeasible requests (4.5x over baseline), preventing hallucinated code.

Surfaced a critical misalignment between heuristic LLM scores (98%) and functional correctness (79% pass), exposing limits of current code assessment.

Launched an AI-powered SaaS platform used by over 30 small and medium-sized businesses, reducing manual expense entry by an estimated 65%.

Led the development and deployment of core AI algorithms for automated expense tracking, real-time analytics, and receipt processing, improving processing speed by 120% over the baseline implementation.

Developed a multi-objective CLIP distillation framework using Soft MoE for per-patch loss weighting (85.3% accuracy, 52.8% mIoU on ImageNet, MEDiC arXiv 2026, ExPLoRe ECCV 2026).

Engineered Cross-Scale MAE, a multi-resolution masked autoencoder for remote sensing achieving 5% accuracy improvement across 4 benchmarks (NeurIPS 2023), with follow-up strategies cutting training time 32% (IEEE IGARSS 2024).

Authored a comparative analysis of LLM code generation security across 4 major models and 9 tasks ("Occasionally Secure," arXiv 2024), uncovering up to 22 vulnerabilities per model and a 40% security regression in Gemini under security-focused prompts.

Led the rapid completion of a critical terabyte-scale data replication project for Greece's leading telecommunications provider, reducing replication time from days to minutes and ensuring real-time views for ETL and analytics.

Spearheaded an ML model to predict order fulfillment times, delivering a 34% reduction over the previous baseline.

Developed a distributed hybrid Girvan-Newman variant on Apache Spark, achieving an 84% runtime reduction on large social graphs without accuracy loss (Algorithms journal, 2019).

Shipped 20+ features for the company's translation CMS, building Python services and SQL-backed workflows used by 300+ translators across 150+ language pairs.