I build across model training, data engineering, Google Cloud, product development and Linux infrastructure. This is the work I can publish.

Some of my professional work is covered by NDAs, so I only show the technical scope I can publish. The product workflows, private data and internal implementation stay private.

I train and serve models

I build labelled datasets, fine-tune language and vision models, compare checkpoints, and turn the selected model into an API that other systems can use.

  • LoRA and supervised fine-tuning, held-out evaluation, error analysis and human review
  • GPU training and inference with batching, async jobs, vLLM and containerized endpoints

I build on Google Cloud

I deploy event-driven services, APIs and data pipelines on GCP, then set up the permissions, secrets, storage and monitoring they need to run safely.

  • Cloud Functions, Cloud Run, Compute Engine, Pub/Sub, Cloud Storage and Artifact Registry
  • Firebase, Firestore, IAM, Secret Manager, BigQuery, Cloud Logging and release automation

I build data systems

I turn messy source data into something models, analysts and products can trust, from ingestion and validation through embeddings, clustering and report generation.

  • Schema validation, entity resolution, quality controls, resumable jobs and human handoffs
  • Embedding services, vector search, clustering experiments and large asynchronous workloads

I run the infrastructure

I set up and operate Linux machines, private networking and protected service access alongside the AI and data workloads running on them.

  • VM and VPS setup, SSH hardening, disabled root access, fail2ban, firewalls and Tailscale
  • Docker, NGINX, TLS, protected API keys, machine scaling, logs, metrics and health checks

Founding Machine Learning Engineer·Apr 2024 to Present

Founding machine learning engineer working across the product’s ML, data and production systems.

3 projectsSee the role

Production ML and Cloud Platform

LLM orchestration, Google Cloud and Linux operations

Built at Intero·Apr 2024 to Present

The platform the machine learning work at Intero runs on. I build the services the product calls, provision and run the machines underneath them, and keep the pipeline between them typed and resumable so a failed stage picks up where it stopped instead of starting over.

  • I build on Cloud Functions, Cloud Run, Compute Engine, IAM and Secret Manager, with Firestore holding application state.
  • I provision the Ubuntu hosts myself, with Tailscale, locked-down SSH, fail2ban, Docker, NGINX, TLS and monitoring.
  • I designed the typed stages, the contracts at each boundary and the event orchestration that moves work between them.

Tools and systems

32B Model Fine-Tuning and Serving

Two QwQ-32B LoRA fine-tunes trained on four H200 GPUs

Private work at Intero·2024 to 2026

The use case and dataset content are private, so this covers the model work only. I fine-tuned two QwQ-32B models for a classification system, from data preparation through training to the served endpoints. The system scored 93% on a benchmark kept out of training.

  • I trained with LoRA at rank 128 and alpha 128 through Axolotl, on 171,031 labelled examples I prepared.
  • Four NVIDIA H200 80GB GPUs per run, four epochs, and I kept epoch two once the later ones started losing ground.
  • I served the selected weights with vLLM on Modal behind OpenAI-compatible endpoints, with FastAPI at the request boundary.

Tools and systems

Read more

Distributed Data and Vision Infrastructure

Async workers, GPU inference and repeatable Linux deployments

Built at Intero·2024 to 2026

The long-running side of Intero. Jobs that chew through a lot of data over hours, on machines that have to recover without anyone watching them. I built the workers, the storage they write into and the GPU vision inference running on top.

  • I built bounded asynchronous workers with batched writes, retries and atomic storage.
  • I run a Dockerized Linux VM fleet with private networking, protected access and repeatable rollouts.
  • I deployed GPU vision inference with versioned outputs, structured logs, metrics and failure recovery.

Tools and systems

Junior Machine Learning Engineer·Jun 2022 to May 2024

Worked remotely with a US team on data ingestion, clustering, machine learning tooling, report automation and production services.

3 projectsSee the role

Clustering and Embedding Infrastructure

GPU embeddings, high-memory clustering and model evaluation

Built at Viable·2023 to 2024

Embeddings want a GPU and clustering wants memory, so I split them into two services instead of one that did neither well. I also built the tooling that decided which models and algorithms we actually shipped.

  • I built a FastAPI embedding service that kept the transformer model loaded on GPU and exposed a narrow authenticated API.
  • I moved long clustering runs into asynchronous Compute Engine workers and stored large results in Cloud Storage.
  • I compared classical, density-based, agglomerative and LLM-assisted approaches with labelled examples and clustering metrics.

Tools and systems

AI-Assisted Data Ingestion

An eighteen-stage pipeline for inconsistent customer exports

Built at Viable·2023 to 2024

Customers sent us files that agreed on nothing. Different columns, different date formats, different identifiers, different encodings. I built the pipeline that turned all of it into one shape the product could load.

  • I separated encoding, headers, dates, identifiers, reshaping and validation into explicit, resumable stages.
  • I used a language model to classify unfamiliar columns and return constrained mappings instead of free-form output.
  • Deterministic checks, dry runs and operator review kept uncertain model decisions away from production writes.

Tools and systems

Report Operations and Quality Systems

Scheduling, parsing and delivery checks for customer reports

Built at Viable·2022 to 2024

Reports went out on a schedule, and most of the work around them was manual and easy to get wrong. I automated the recurring parts and put checks in front of delivery so a bad report did not reach a customer.

  • I built a scheduled Cloud Function that handled report cadence, company resolution and operational updates across several services.
  • I used confidence thresholds and deterministic rules to automate ordinary cases while keeping uncertain matches visible to operators.
  • I built parsers that checked citations, quotations, empty sections and low-support findings before delivery.

Tools and systems

Independent software engineer·2026 to Present

Websites and internal systems I designed and built for businesses, with each product, deployment and data boundary kept separate.

1 projectSee the role
01 / 03

Vipulse production home page

Vipulse Corporate Website

A multilingual website and publishing system for electronics manufacturing

Client project·2026

I designed and built Vipulse's production website, publishing platform and lead-capture system. The Next.js application serves separate English and Hungarian content from Payload CMS and PostgreSQL, with technical product pages, structured metadata, consent-aware analytics, spam controls and a preview-to-production release workflow.

  • A Next.js and Payload CMS platform with separate English and Hungarian content, previews and structured publishing workflows.
  • Technical service pages, reusable product modules, metadata, sitemaps and schema markup built for industrial buyers and search engines.
  • Protected lead capture, rate limiting, transactional email, consent-aware analytics and controlled Vercel releases.

Tools and systems

Read more

BSc Computer Engineering·Sep 2019 to Jun 2023

Computer engineering fundamentals, applied software and machine learning projects, and graduation as valedictorian.

1 projectSee the role

Tweet Noise Classifier

Training an ALBERT model to separate signal from noise on Twitter

Built at Manisa Celal Bayar University·Dec 2022 to Jan 2023

For my graduation project, I trained an ALBERT-xxlarge-v2 classifier to separate useful tweets from bots, marketing and unrelated posts around a trending topic. I built the collection pipeline, a human annotation interface, Firestore persistence, the PyTorch training and checkpoint flow, and a Flask API that served the selected model for batch classification.

  • I collected topic streams, defined signal and noise labels, and built a Flet interface for reviewing and annotating tweets.
  • I fine-tuned ALBERT-xxlarge-v2 with PyTorch and Hugging Face, then loaded the selected checkpoint for inference.
  • Firestore connected collection, annotation and batch classification; Flask exposed the model through a small HTTP service.

Tools and systems

Read more

Nights and weekends·Ongoing

Open-source experiments and personal tools I build to learn, solve a problem or explore an engineering idea.

1 project
01 / 03

The store banner. Same circular link mark as the toolbar icon, on a transparent canvas.

Copy URL Shortcut

⌘⇧C copies this tab's URL. Nothing else runs.

Built for myself·Aug 2026

Chrome has no built-in shortcut for copying the address of the tab you are looking at. Arc did (Copy Page URL on ⌘⇧C), and the store listings that come close either inject a content script into every page or rewrite the copied string. This extension copies location.href verbatim on that chord, or on a toolbar click, then goes idle. No host permissions, no network, no analytics. Live on the Chrome Web Store from day one.

  • Manifest V3 uses activeTab only, with no host permissions, network requests, analytics or background page access.
  • A closed-shadow toast confirms the copy without letting the current website restyle the extension UI.
  • The public repository includes manifest, clipboard and page-context tests plus the Chrome Web Store package workflow.

Tools and systems

Read more