ECE student at Aristotle University of Thessaloniki. I build AI agent tooling, parallel systems in C/CUDA, and full-stack apps in Next.js and TypeScript.
How I work: I work with coding agents and review/verify everything that lands —
see fuzz/README.md
for how I validate generated code against sanitizers.
Live demos: CookBook · NexusFit
AI agents
- 🔁 contributor-sandbox — delegation pipeline for cheap coding models: measured a 15M-token exploration cost in a real run, designed the prompt/sandbox process that removes it, and shipped the safety checklist + cost-analysis tooling.
- 🌉 dsh-agent-bridge — run a coding agent as a tool-callable subagent over the Agent Client Protocol, with a deterministic verification gate that stamps unverified output instead of trusting it.
Full-stack & frontend
- 🍽️ OrderReady — multi-tenant QR ordering + live kitchen dashboard for restaurants (Java 21 / Spring Boot, PostgreSQL, Docker): signed table-session tokens, server-side price authority, Greek VAT engine and a transactional outbox to thermal printers. Technical showcase, source available on request.
- 🏋️ NexusFit —
my own gym-platform product in Next.js 15 / React 19: marketing site, member portal,
admin CMS and an AI workout copilot. Multi-tenant over SQLite with offline/PWA mode, rate
limiting and signed sessions; one config object drives the whole deployment, with Docker
- Caddy and 9 Vitest suites. AGPL-3.0. Try the demo →
- 🍳 CookBook — recipe + culinary-jobs platform: React 18 / TypeScript / Vite 6 front end over an Express 5 + SQLite API, responsive from 375 px phones to desktop. The UI doubles as a spec-conformance harness — every Gherkin acceptance scenario from the SRS can be driven into its exact state live. Try the demo →
Systems & performance
- ⚡ cuda-bfs-afforest — parallel BFS & Afforest connected components in CUDA: direction-optimizing traversal, delta-varint compression, 3.6B-edge Friendster graph runs.
- 🧵 mpi-opencilk-connected-components — hybrid distributed/shared-memory connected components: MPI across nodes, OpenCilk within each process, built with PGO and LTO.
- 📡 bluesky-firehose-collector — multithreaded real-time collector in C for the Bluesky Jetstream firehose on a Raspberry Pi Zero W (ARMv6). 86,394 of 86,400 one-second samples, p99 deadline jitter 845 µs, zero dropped frames across 3.4M messages — plus the analysis and the report.
- 📶 satellite-receiver-sim — 20 GHz DVB-S2X 64-APSK receiver simulator: RF front-end cascade analysis, ITU-R propagation models and dual-path stage-by-stage simulation in C, against a MATLAB reference.
Applied ML
- 📈 drl-trading-agents — deep RL trading agents on Gymnasium/gym-anytrading (A2C baseline, modular pipeline).
- 🧠 cifar10-conv-autoencoder — ResNet-style convolutional autoencoder for CIFAR-10 with masked learning and a joint MSE+SSIM loss.
- Building: a Kubernetes/Docker platform project — containerized ML inference served on a cluster with CI/CD, GitOps, and autoscaling.
- Open for entry-level roles where agents do the work and I make them reliable.