Internal
Designed and optimised a self-hosted ~27B-parameter LLM inference platform supporting multi-user workloads.
Model serving, quantization and precision trade-offs, GPU memory, request queues and concurrency, with monitoring and load testing under simulated agent workloads.
Stack: Python, Docker, GPU inference servers, Linux
Maritime Digitalisation Centre
Full-stack platform for real-time vessel tracking from multiple AIS sources, built across the stack from geospatial storage to the live map.
Historical vessel analysis, geofencing and areas of interest, and vessel activity visualisation on continuously generated real-world data.
Stack: React, Node.js, PostgreSQL/PostGIS, Redis, WebSockets
Internal
Designed and deployed an internal containerised application platform supporting multiple production research applications.
Isolated workloads, private networks and restricted secrets. A configuration watcher validates routing changes before applying them, so a bad deploy cannot break production.
Stack: Docker, Linux, Nginx, Cloudflare, CI/CD
Research and technical documents
OCR, information extraction and LLM-assisted analysis for technical and research documents.
Structured data pipelines and APIs that feed the results into larger research and software systems.
Stack: Python, OCR, LLM APIs, PostgreSQL
Live product. Founder and engineer
An AI-assisted meal-planning product that I designed, built and operate, from frontend and backend to databases, AI integration and production infrastructure.
It plans meals for the week, organises recipes, tracks calories and macros automatically, and turns the plan into a grocery list.
Stack: Next.js, Prisma, Auth0, Stripe, Sentry, PWA
mealverve.comVoice AI and tool calling
An English and Greek voice agent for outbound and inbound calls. It checks real calendar availability, books the meeting and texts a confirmation with an add-to-calendar link.
Ran five concurrent outbound calls. Detects voicemail and IVR systems and ends the call instead of talking to a machine.
Stack: Vapi, GPT-4o, ElevenLabs, Deepgram, Twilio, Telnyx, Flask
Multi-agent and realtime voice
Speak a task into a phone app. A router model sends it to Claude Code, Codex or a Qwen coding agent, which works in an isolated git worktree. JARVIS reports back by voice and merges on command.
Agent sessions resume per project. When an executor hits a rate limit it goes on cooldown and the router picks another.
Stack: Node.js, WebSockets, Flutter, Qwen realtime voice, Claude Code, Codex
Autonomous plan, code and review loop
Planner and reviewer agents direct coding agents through a task checklist. Independent subtasks run in parallel in separate git worktrees and merge back.
The reviewer checks UI changes in the running app through Playwright MCP and database state through a read-only database MCP. I used it to build a full-stack compliance web app.
Stack: Python, Claude, Codex, MCP, Playwright, git worktrees
Document AI, Zindi competition
Handwritten-text recognition for historic archival documents, with a 5-fold evaluation harness and a data audit before any training.
Stack: Python, PyTorch, Kraken, McCATMuS
B.Eng. thesis, University of Cyprus, 2023
Lead programmer in a five-person team. LSTM networks forecast blood glucose and a reinforcement-learning model calculates insulin doses, both running on Android through TensorFlow Lite.
Stack: Python, TensorFlow, Keras, TensorFlow Lite, Android
Code on GitHubI'm most useful when the problem is complex or not yet well defined.
Agents that do real tasks, connected to your databases, CRMs and internal APIs.
Search and question answering over your company's documents and data.
OCR and structured extraction, fed into the systems you already use.
Making an AI prototype reliable: evaluation, monitoring and cost control.
Debugging and improving RAG, agent and LLM pipelines that aren't performing.
Running models on your own hardware, and routing between models to balance cost and quality.
2023 – present
Cyprus Marine and Maritime Institute, Maritime Digitalisation Centre
2019–2023
University of Cyprus
R&D: an Android phone captures colour, depth and camera pose, and a Python pipeline turns it into a metric point cloud and mesh. Kotlin, Rust, PyTorch, depth models.
Transcribes whatever a PC is playing, word by word, on the CPU, with Moonshine and Whisper.
The first things I built with software and AI.
Builds playlists from a plain-language description, with OpenAI and Stripe.
Booking site for a personal trainer. Bookings go to Google Sheets with email notifications.
Translation tool built on the OpenAI API.
Android app for habit tracking and tasks, with a Firebase backend.
Simulator of a pipelined RISC processor, in C++.
Smart rockets that learn to reach a target, in p5.js.
Retail point of sale software, in C++ with Qt.
Email is the quickest way to reach me.
charalambosrotsides@gmail.com