AI Engineer, LLM Engineer, AI Solutions Architect, Forward Deployed Engineer

Senior individual contributor. Machine learning from 2017, agent systems from 2023.

  • Athens, Greece
  • Remote, EU and global
  • EU work authorisation
  • Greek, English
  • Open to contract and permanent

Start a conversation LinkedIn

I did not move into AI. The field arrived where I already was: forecasting, pricing and segmentation running against live commercial data from 2017, five years before the generative wave, inside jobs titled Full Stack Developer.

Technical skills, the year each first shipped, ownership, and a link to the evidence.
Skill Since Role Built with it
LLM and agent systems 2023 owned Production agent systems with evals and cost ceilings
Evaluation and reliability 2023 owned Agent harness benchmark, 45 graded sessions, data published
Cost and latency engineering 2023 owned Seven versions of one agent, including the two that got worse
Language model training from scratch 2026 owned BSLM, a 72M parameter model, public repo and weights external
Retrieval and agent memory 2023 owned Ten running systems, each with its own retrieval and memory layer
Demand forecasting and algorithmic pricing 2017 owned Powercore: forecasting and automated repricing on a live catalogue
Customer segmentation and revenue forecasting 2019 owned Developgreece: models live in client marketing workflows
Data engineering 2017 owned Ingestion, schema design, validation and quality control
Web platform and performance engineering 2014 owned 25 live sites, custom ERP and CRM, Lighthouse 96 to 100
Delivery inside a cross-functional team 2017 in a team Two employed roles: requirements and delivery with client teams

Also used: LangChain, LangGraph, RAG, embeddings, hybrid search, reranking, MCP, tool calling, structured output, vector stores, Python, FastAPI, PHP, JavaScript, MySQL, PostgreSQL, Redis, Docker, Git, REST, SEO, CRO, analytics.

Where it happened

  • Powercore Sept 2017 to Oct 2018 Full Stack Developer and Marketing Advisor, employed. Forecasting, pricing and the data layer under both.
  • Spoilsins.com Nov 2018 to present Founder and sole engineer of a content platform. Built twice, eight years apart.
  • Developgreece May 2019 to Sept 2022 Web Developer and Project Manager, employed. Agency delivery plus custom models for client platforms.
  • Broikos Oct 2020 to present Independent. Retrieval, agent orchestration, cost and latency work, AI adoption advisory.
  • Parthand Mar 2021 to June 2023 Co-owner of a handmade womenswear e-shop. Built it, ran it, grew it.

Three of them, in full

Pricing a catalogue against a market that moves daily

Powercore, 2017. The interesting part is not the model, it is the closed loop: collect competitor prices on a schedule, normalise them, match them to the catalogue, forecast demand, set a price against a margin constraint, measure, repeat. Nobody in the middle. The OpenCart development was two or three days a month; the rest of the year was spent building the thing that decided what the prices should be. In 2017 there was no title for that work, so it sat inside a job called Full Stack Developer.

The same product, built twice, eight years apart

Spoilsins, from 2018. The first version ran an automated content pipeline on the natural language processing of its day: extractive summarisation, TextRank and TF-IDF scoring, candidate selection, and the corpus engineering underneath it. The second version, after generative models became production technology, was a re-architecture around autonomous content and quality assurance pipelines that took the operational workload to near zero. What carried over was not the model. It was the data layer and the evaluation.

Segmentation a marketing team actually used

Developgreece, 2019 to 2022. Custom models for client e-commerce platforms: customer segmentation, demand and revenue forecasting. Problem framing with the client, feature engineering, training data, algorithm selection, evaluation, then deployment into somebody else's marketing and merchandising workflow with monitoring and retraining after release. Agencies are where models go to die, because they have to survive a team's existing habits. These shipped and stayed shipped.

How I work

  • You brief the person who builds it. Nobody in between to misunderstand you.
  • Every build answers to a number. If none moves, it was a cost.
  • Evaluation before launch, not after. A system nobody measured is a system nobody can fix.
  • Attribution always. Agency work is credited to the agency.