AI Systems & Agents
LLM-powered products, assistants and multi-agent workflows with tools, memory and guardrails. RAG over your documents and data, with retrieval you can measure — in the cloud or fully on-premise.
[ AI · ML · Full Stack ]
I design AI systems, train and fine-tune models, and ship the software around them — from data pipeline to the app your team actually uses.
Local time Iași --:--
01 Introduction
AI engineer, full stack by trade
And everything it needs to survive contact with real users. For more than a decade I've shipped production software — Java and Spring backends, Angular, React and Vue frontends, native and hybrid mobile apps. Today that experience goes into AI: LLM-powered products, voice and document AI, agents, fine-tuned models and automations that remove hours of manual work every week — much of it running privately, on the client's own GPU servers.
I care about the boring parts that make AI useful: clean data, honest evaluation, sensible costs and software that someone can still maintain years from now. An application should be built to last a decade, not a demo day.
02 What I do
From idea to production, end to end
One person who understands the model, the backend and the interface means fewer hand-offs and fewer things lost between them. Pick one piece or the whole chain.
LLM-powered products, assistants and multi-agent workflows with tools, memory and guardrails. RAG over your documents and data, with retrieval you can measure — in the cloud or fully on-premise.
Dataset curation, LoRA / QLoRA fine-tuning, embedding and reranker models, and evaluation suites — so a smaller model can do your task better and cheaper.
Document AI and OCR, classification, prediction, image and voice understanding — built around your data, with interpretable outputs and a deployment path that fits your infrastructure.
Repetitive processes turned into pipelines: data extraction, reporting, integrations between the tools you already use, and AI steps where rules aren't enough.
Web platforms, internal tools and mobile apps — Java / Spring Boot or Node backends, Angular, React or Vue frontends, Android and iOS clients. Built to last a decade.
03 Selected work
Shipped, maintained, still running
URBIO's smart city platform — remote monitoring, scheduling and control of public lighting, smart poles and city infrastructure on a live digital-twin map. I work across the stack on the new NEOS interface, lighting programs, the digital twin and smart-pole controls, and bring AI into the platform and the teams around it.
A private AI platform on dedicated GPU hardware: large language models, real-time voice-to-voice conversation with natural turn-taking and voice cloning, and a personal agent that acts across your tools and documents. Romanian-first, with no data leaving the platform.
The archive that knows what it contains. Nabu reads scanned documents, catalogues them, masks personal data (CNP, IBAN, addresses) on text and page image, and answers questions with a citation to the exact page. It runs entirely on a server inside the client's network — no cloud, no per-page costs, around 1,800 pages an hour.
nabusoft.ro →
Energy and gas bills, brought in automatically from supplier portals, RO e-Factura or email and checked line by line. Tracks surplus kWh, compares the bill with inverter and smart-meter data across ten inverter brands, and generates documented complaints — plus a white-label portal for solar installers.
e-asist.ro →
Interactive roadmaps for developers in Romanian — learning paths by role, guides, progress tracking and an AI tutor running on a self-hosted LLM. Rebuilt from scratch with Google sign-in, an admin dashboard, Pro plans and AI credits.
reprofil.ro →An AI education platform that runs on the school's own hardware. RAG over course materials in Romanian, AI-drafted lessons and assessments that the teacher reviews and approves, GDPR tooling, rate limiting and full observability. The AI assists — the teacher decides.
A configurable, versioned workflow engine for public tenders: it reads the tender documentation, extracts every requirement with its evidence, matches it against the company's people and resources, fills the forms and assembles the submission archive.
n8n pipelines on top of Jira that read scanned documents with vision LLMs, classify and validate them, track deliveries, and assemble complete technical construction files and official reports per site — dozens of documents per file, filled in and filed without manual copy-paste.
An assistant trained on your documents, a model fine-tuned for your task, or the automation that frees your team from copy-paste work. Tell me what slows you down.
Start a conversation →+ More work
Enterprise multi-agent assistant: six domain agents, a Jira gateway with per-profile access policy and company single sign-on.
Agents · Dify · SSOMulti-agent research automation — eight specialised agents, document ingestion, RAG over a vector store and an async execution pipeline.
Spring Boot · Qdrant · RabbitMQA frontier open-weight model with a 512k-token context served across two DGX Spark nodes over RoCE, behind an OpenAI-compatible endpoint.
vLLM · Ray · RoCEScanned PDFs, images and Office files to structured Markdown — tables, formulas, reading order — on GPU, shipped as air-gapped bundles for ARM64 and x86.
MinerU · vLLM · DockerSelf-hosted neural web search: crawling, hybrid BM25 and vector retrieval, and reranking — all on local models, no paid APIs.
Java · OpenSearch · QdrantControl plane for LLM inference: agent registry, model-aware routing, streaming passthrough, circuit breakers and per-client usage accounting.
WebFlux · PostgreSQL · AngularMission-control dashboard for orchestrating an AI agent fleet, paired with a self-hosted knowledge vault, shipped as a portable Docker bundle.
Next.js · Framer MotionAndroid phone automation driven by a vision-language model — the agent sees the screen and taps, types and swipes through ADB.
Spring Boot · ADB · VLMCommand and telecommunication layer for outdoor devices over a self-healing WiFi mesh, with a gateway, OTA updates and secure provisioning.
IoT · Mesh · EmbeddedTwo lead management services for Harte Hanks, plus data services that keep multiple SQL servers in real-time sync.
Java · Spring Boot · VueConstruction management software that runs an entire business on web, Android and iOS, and smart digital inspection forms.
Demo · CheckListModules for a satellite client interface on Eclipse RCP and OSGi for Akka Technologies, including the Java 6 → 8 migration.
OSGi · RCP · Angular04 How I work
Most AI projects succeed or fail on the data. I look at it first — then decide if you need a model, a prompt, or just a query.
Every model change is checked against an evaluation set built from your real cases. "Seems better" isn't a metric.
A working slice in front of users beats a perfect plan. Short iterations, visible progress, no surprises.
Readable code, sensible architecture, documented decisions. The person maintaining it in five years should thank us.
05 Toolkit
Twelve years in, working with
Not a badge collection — these are the tools I know well enough to tell you where they save a week and where they quietly cost one.
Sector_01AI & Machine Learning
Sector_02LLM & Agents
Sector_03Backend
Sector_04Frontend & Mobile
Sector_05Data & Automation
06 Career log
URBIO · Iași
Harte Hanks · Iași
MagiTech Construction Software · Remote
Akka Technologies · Iași
InnoDevs · Iași
Web & mobile apps for small and mid-size clients
[ Transmission_form ]
Tell me what you want to build, your timeline, and anything else I should know. I usually reply within 24 hours.