Open to remote roles across Europe

Reliable backend systems.Local-first AI tools.

Software developer. PostgreSQL-backed APIs, reliable event delivery and transactional inventory, tested under failure and contention. Also retrieval, agent workflows and local inference.

Selected work / 2026

Reliability under failure. Correctness under contention.

Two systems. RelayForge keeps an event moving after a receiver breaks. Slotline keeps a seat from being sold twice when every buyer clicks at once.

RelayForge Control Room showing measured delivery KPIs, the pipeline, a live event stream and system healthReal capture · Control Room
Independent buildBackend infrastructure · Distributed systems

Case study · Durable event delivery

RelayForge

A self-hostable event pipeline: PostgreSQL as the queue, lease-fenced workers, signed retries, dead letters and audited replay. The Chaos Lab breaks a real receiver and lets you watch the backend recover.

  • PostgreSQL
  • Fastify
  • TypeScript
  • SSE
System
PostgreSQL queue · SKIP LOCKED
Behavior
At-least-once · fenced leases
Evidence
56 tests · 6 E2E · local benchmark
Slotline storefront with ceramics studio photography, live session availability and a reservation panelReal capture · Storefront
Independent buildFull-stack SaaS · Transactional commerce

Case study · Reservation commerce

Slotline

A customer booking product and merchant console built around one invariant: held plus confirmed spots can never exceed capacity. Holds expire on the database clock, checkout confirms under the same lock, and the Contention Lab measures the race.

  • PostgreSQL
  • Fastify
  • React
  • Payments sandbox
System
PostgreSQL inventory · Fastify API · React console
Behavior
held + confirmed ≤ capacity
Evidence
100 requests → 3 holds · 0 errors
49 tests · 9 E2E
Independent buildPrivate working app · 2026

Case study · Local-first writing

Private AI Writer

Android writing app + local desktop backend. Markdown drafting, AI assistant with tool-based edits, no cloud model.

  • React
  • TypeScript
  • Capacitor / Android
  • Node.js / Express
  • llama.cpp
  • Vitest
Concept illustration of a phone document with a proposed revision, diff and undo control
Concept illustration Revision diff and undo
System
Android client · Express API · llama.cpp
Behavior
Guided edits · revision checks · diffs · undo
Evidence
3 real UI captures · deterministic tests
How it works

Architecture.

React/Capacitor → authenticated Express API (LAN) → llama.cpp server. Markdown library, revision history.

Edit safety.

Section-level reads, structured questions, targeted patches. Stale-revision rejection; diffs and undo in history.

Testing.

Streaming tool/edit progress; resumable approvals. Deterministic tests: patching, cancellation, stale revisions, undo.

Android clientLocal Express APIllama.cpp inference

Markdown files + revision history

Independent buildPrivate working app · Local generation, 2026

Case study · Local image generation

AnimaHub

Mobile-first PWA for local image generation. Prompt builder, presets, composition controls; Python service runs generation on a desktop GPU.

  • Python
  • JavaScript
  • PWA
  • stable-diffusion.cpp
  • Vulkan
System
PWA · Python HTTP service · stable-diffusion.cpp
Behavior
Prompt building · presets · generation locking · cancellation
Evidence
2 real UI captures · health checks and cancellation
How it works

Interface.

Grouped prompt sections, tags, presets, touch controls for identity, appearance, composition, rendering.

Backend.

Local HTTP service: model process management, generation requests, presets, gallery output, per-model config.

GPU constraints.

Vulkan on an 8 GB AMD GPU; tiled VAE decoding, CPU text encoding; LoRA and subprocess fixes.

Mobile web / PWAPython HTTP backendstable-diffusion.cpp

Local GPU generation via Vulkan

Systems & agent tooling

Deterministic state, typed actions, revision-safe writes.

Engine owns state · model writes prose

Typed actionDeterministic stateNarration

Context compiler · provider contract · evaluation harness

System boundary · state stays outside the model
Personal R&D prototypeTypeScript · Node.js

Narrative engine · context assembly · evaluation

Matrix 2.0

Typed actions and deterministic rules own state; the model writes narration from a bounded context packet.

  • Deterministic state
  • Context compilation
  • LLM contracts
  • Evaluation
Read more

Typed actions → deterministic resolution → state patch + event log. Bounded context packet from current state (no transcript). Batch harness: scorecards, latency, prompt size. Contract + pipeline tests.

Systems Map overview showing an example map with colour-coded concept states and relationships
Real UI capture · working prototype
Working prototypeNode.js · Agent CLI

Structured knowledge · revision-safe writes

Systems Map

Shared structured model for a person and a coding agent. Revision tokens, stale-write rejection, all-or-nothing batches.

  • Structured agent CLI
  • Revision-safe writes
  • Atomic batches
  • Local-first
Read more

Shared structured model for a person and a coding agent. Agent CLI: revision tokens, stale writes rejected, batches up to 100 commit all-or-nothing. Deterministic serialization; ideas / accepted / verified separate.

Public benchmarkEvaluation · HTML5 Canvas

Empirical evaluation · multi-model comparison

Aquarium LLM Benchmark

A controlled comparison of LLM-generated game code for one canvas task, with a public dashboard that shows each result and its diagnostics.

  • LLM evaluation
  • Canvas
  • Diagnostics

Featured demo / Voice + agents

Voice-Enabled Pi Coding Agent

Working demo

Push-to-talk CLI agent voice layer. Hold key → ffmpeg → persistent faster-whisper. Final answer → local GGUF TTS → 24 kHz PCM streamed (speech before synthesis finishes).

PCM envelope drives an Electron avatar over WebSocket; recording cancels playback. Three TypeScript extensions + overlay, one machine.

  • TypeScript
  • Electron
  • WebSockets
  • faster-whisper
  • GGUF TTS
Read more
Real capture 18 seconds · Sound on · Tool activity and live speech

Further exploration / Local inference

Local LLM Steering Lab

Inference experiment

llama.cpp / AMD Vulkan activation-steering playground. Python CLI, browser UI, control-vector extraction, strength/layer controls, live steering without reload.

Serialized inference, streamed responses, baseline comparisons, repeatable experiments.

  • Python
  • llama.cpp / C++ integration
  • GGUF control vectors
  • Vulkan

Inspired by LynnColeArt/ai-hotbox ; built using llama.cpp’s control-vector machinery.

Context Nexus running inside Unreal Editor with Project Explorer, static analysis, and AI assistant panels
Real product capture · Unreal Editor integration
Archived R&D

Developer tooling / Unreal Engine 5

Context Nexus

Unreal Editor assistant: Blueprint DSL, Clang 18 C++ analysis, focused LLM context.

  • UE5
  • C++
  • Clang 18
  • LLM APIs
Visit project site

Archived after native Unreal AI and MCP tooling covered similar work.

Games & interactive

Selected builds · Prototypes · Game jams

Evolution War · Apartment Panic · Orange Clicker · Escapelike · Project Y · UE5 devlog · 2 jam games.

  • Orange ClickerFinal project
  • Evolution WarFinished release candidate
  • Apartment Panic3D prototype
  • EscapelikePlayable prototype
  • Project YMultiplayer prototype
  • Project EmberUE5 devlog
  • Tough CrowdGame jam
  • 19,999 GoldGame jam

02 / Experience

Independent since 2025. Level 2 escalation, 2023–2025.

Jul 2025 – Present

Independent

Backend & applied AI · Remote

Software developer

Backend systems, applied AI and retrieval

  • Built RelayForge and Slotline: PostgreSQL-backed event delivery and booking systems with concurrency tests, failure injection and Playwright browser flows.
  • Private AI Writer, AnimaHub and a voice coding agent: UI, API and local model integration.
  • Retrieval and agent systems: chunking, local embeddings, hybrid retrieval, reranking, validated tool calls, retrieval-quality evaluation.
  • Local inference on llama.cpp / GGUF: quantization, GPU/CPU offload, KV cache, activation steering (AMD Vulkan).
Jan 2023 – Jul 2025

Teleperformance

Automotive programs · Greece

Expert Team Specialist / Technical Support Specialist

Audi digital services & Mercedes-Benz roadside assistance

  • Promoted to Expert Team 2023; about 30 Level 2 escalations per 8-hour shift.
  • Salesforce + Audi internal diagnostic tools: backend log analysis, system resets, cached service data invalidation.
  • Updated infotainment and remote-control service packages; step-by-step troubleshooting for dealers and phone agents.
  • Improved dialogue trees, procedures, technical documentation.
  • Mercedes-Benz roadside assistance support.

03 / Capabilities

Backend, retrieval, agents, inference.

01

Backend systems

PostgreSQL-backed APIs · reliable delivery · transactional inventory · tests

  • PostgreSQL
  • Row-level locking
  • Idempotency
  • Lease fencing
  • Fastify
  • SSE
  • Playwright
  • TypeScript
02

Applied AI

Local inference · retrieval · agent workflows · evaluation

  • llama.cpp
  • GGUF
  • faster-whisper
  • Hybrid search
  • Reranking
  • Tool calling
03

Systems & operations

Interfaces · real-time systems · technical escalation · documentation

  • React
  • Node.js
  • Python
  • C++
  • WebSockets
  • Cloudflare
  • Git
  • Level 2 escalation
  • Salesforce

04

Contact

Roles: backend engineering · full-stack · applied AI · automation · solutions · technical support engineering. Remote, Europe. Open to relocation.