Panagiotis Doumas

Project Archive

Selected work across AI, software, and research.

A collection of client work, applied AI systems, and university research—from factory optimisation and LLM compression to robotics, IoT, and smart contracts.

← Back to selected work

All projects

Showing all projects.

2026

SEO Automation · Generative Engine Optimisation

Automated SEO Content System

Designed and built an n8n workflow that replaced a manual content-production loop with a repeatable route from keyword opportunity to a published, search-optimised page.

AutomationOrchestrated keyword research, search-intent analysis, content generation, quality checks, on-page optimisation, and direct website publication in one workflow.
SEO + GEOPaired technical performance with clear, structured topical coverage so the pages could serve conventional search journeys as well as AI-assisted discovery.
Measured outcomeThe workflow supported top-five rankings across 20+ keywords, stronger CTR and engagement, 95+ Lighthouse scores, and visibility in relevant AI-search and chatbot results.
n8nSEOGEOAI AutomationTechnical SEOContent Strategy
2024

AI / ML Research · Master’s Thesis

LLMZip — Data Compression Using Large Language Models

Master’s thesis conducted under academic supervision, investigating how large language models can improve text compression.

MethodDesigned and implemented a hybrid compression algorithm that leverages LLMs to model text and achieve performance beyond traditional methods.
ResearchConducted novel cross-linguistic entropy analysis and established the lowest entropy lower bounds ever reported for multiple languages, including English.
OutcomeOutperformed conventional compression approaches and published the findings in a dedicated technical report.
PythonPyTorchCUDALarge Language ModelsData Compression
GitHub ↗
2023

AI Optimisation · Client Collaboration

Adaptive Scheduler for Manufacturing Plant

Developed with CloudNC, this proof of concept explored how real-time factory data could drive more responsive production planning.

Signal to scheduleInterpreted CNC-machine signals alongside machine availability and operating constraints, then used AI optimisation to assign and reassign production jobs.
Adaptive behaviourThe planner re-optimised when a real-time event changed the factory state, rather than relying on a fixed, one-time schedule.
Client deliveryValidated the potential for optimisation algorithms to improve factory-output planning while aligning the proof of concept with CloudNC’s specifications for production-ready software.
CNC Signal AnalysisAI OptimisationProduction PlanningManufacturing SystemsClient Delivery
2023

Robotics · University Project

Robotic Manipulation

Programmed a four-degree-of-freedom robotic arm in MATLAB to complete increasingly complex manipulation tasks with a custom-designed gripper.

ManipulationControlled two arm joints and a gripper to pick and place coloured cubes at a target position while orienting the red face upward.
Mechanical designDesigned and 3D-printed a CAD gripper for drawing a dimensioned triangle with a marker, then reused its dimensions to create a knife handle for a fruit-cutting task.
CollaborationCombined motion-control work, mechanical prototyping, and clear team communication to deliver all three tasks safely.
MATLABRoboticsCAD3D PrintingMotion Control
2023

IoT · Full-Stack Development

SeatSense

Built an IoT web application that shows live seat availability across a university library, addressing the difficulty of finding a space during peak hours.

DeviceUsed a Raspberry Pi and temperature sensor to detect whether a seat was occupied, including conversion of raw readings into Celsius values.
Data flowSent sensor updates to Firebase, where the backend retrieved them for a full-stack interface with live, floor-by-floor availability maps.
Product focusTurned a shared student frustration into a practical prototype by connecting accurate sensing with an immediately understandable interface.
CRaspberry PiFirebaseIoTFull-Stack Development
2023

AI / ML · Computer Vision

MRI Segmentation with U-Net

Implemented a U-Net in Python and PyTorch from the original architecture to segment MRI imagery against provided ground-truth masks.

ArchitectureBuilt the encoder-decoder network and observed how skip connections preserve spatial information for more accurate image segmentation.
EvaluationMeasured segmentation quality with F1 score and Jaccard Index, obtaining predictions that closely matched the ground truth.
LearningStrengthened practical PyTorch experience while translating a research-paper architecture into a working implementation.
PythonPyTorchU-NetComputer VisionMedical Imaging
2023

Web3 · Smart Contracts

NFT Ticket Marketplace

Developed an end-to-end Solidity marketplace for event-ticket NFTs, covering token creation, primary sales, and resale through an escrow-based secondary market.

ContractsImplemented an ERC-20 payment token and ERC-721 ticket NFTs, with ticket metadata for event details, ownership, validity, supply, and use status.
MarketsCreated primary ticket sales and a secondary marketplace where owners list tickets, buyers submit bids, and creators receive a 5% royalty on accepted resales.
SecurityUsed Solidity validation with require statements and extensive Forge tests to verify contract behaviour and reduce vulnerabilities around external calls.
SolidityERC-20ERC-721FoundrySmart Contracts
GitHub ↗
2024

AI / ML · Generative Models

Generative Models: VAE and DCGAN

Built and trained a variational autoencoder on MNIST and a deep-convolutional GAN on CIFAR-10, exploring the practical tradeoffs of generative-model design.

VAEImplemented convolutional encoder and decoder networks, the reparameterization trick, and a beta-VAE loss while studying latent-space interpolation and disentanglement.
ExperimentsTested beta values to understand posterior collapse and how the regularization balance affects the usefulness of latent representations.
DCGANIterated on generator and discriminator architectures, learning rates, epochs, and latent-vector size to produce recognizable CIFAR-10-like images while avoiding mode collapse.
PythonPyTorchVAEDCGANMNISTCIFAR-10
2023

Multi-Agent Systems · Collaborative Engineering

Self-Organising Multi Agent Systems Simulation

Co-built a Go-based multi-agent game environment and a competitive agent for a large cohort simulation involving megabikes, lootboxes, governance systems, and environmental hazards.

Shared worldHelped create the full SOMAS environment—the game world in which agents live, form groups on megabikes, collect lootboxes, navigate hazards, and make governance decisions.
Agent strategyDesigned an agent that prioritised survival and nearby energy sources before pursuing same-colour lootboxes, with an Honesty Matrix to reward reliable agents and penalise deceptive behaviour.
OutcomeOur agent ranked among the top performers across thousands of simulation runs, while the shared codebase developed experience in large-scale Git collaboration and inter-team coordination.
GoMulti-Agent SystemsAgent-Based SimulationGame TheoryGitHub
GitHub ↗
2025

Applied AI · Financial Data

Stock Sentiment and Algorithmic Trading

Created a stock-news sentiment scraper and an algorithmic trading bot that used live RSS feeds and current market data to inform trading decisions.

News ingestionCollected live RSS updates from trusted news sources and transformed article coverage into stock-specific sentiment signals.
Market integrationConnected the trading component to live market data so decisions could react to current conditions rather than static datasets.
FocusExplored the full path from unstructured financial news to a quantitative signal and an automated trading workflow.
RSSSentiment AnalysisMarket DataAlgorithmic TradingAutomation
2025

Web3 · Quantitative Trading

Hyperliquid Perpetuals Trading Bot

Built a Python quantitative trading bot for Hyperliquid that uses real-time price data from perpetual-contract markets to drive automated trading decisions.

Real-time dataConsumed live perpetual-contract price data so the strategy could respond to current market movements rather than delayed or static inputs.
Quantitative logicApplied algorithmic trading logic to the incoming market data to generate and act on systematic trading signals.
Market focusExplored the engineering requirements of automated execution in crypto perpetuals, where continuous price updates and disciplined data handling are essential.
PythonHyperliquidPerpetual ContractsReal-Time Market DataQuantitative Trading