UK. Let’s talk ↗

AI Engineer · Netherlands

I build LLM systems—
and the evals that keep them honest.

Ten years in software engineering, from frontend leadership to production AI. I focus on agentic systems, deterministic quality gates, and local-first infrastructure.

View selected work ↓ GitHub ↗ LinkedIn ↗

01 / SELECTED WORK

Systems that show
their work.

01

LOCAL-FIRST AI · OPEN SOURCE

GitHub ↗

Samuraizer

Private meeting intelligence, entirely on your machine.

Turns recordings into transcripts, summaries, action items, and decisions without sending data to the cloud. The reference implementation of the memnex v0.2 open specification, with portable output and a complete provenance chain.

PROBLEM

Useful meeting memory without cloud exposure or vendor lock-in.

SYSTEM

Resumable CLI pipeline, parallel LLM stages, and an MCP interface for agents.

RESULT

Published as two npm packages; ~2.3× faster analysis through parallel execution.

  • TypeScript
  • Whisper
  • Ollama
  • MCP
  • Node.js
02

AGENTIC PIPELINE · OBSERVABILITY

GitHub ↗

Pixelize

A photo-to-pixel-art agent that exposes every decision.

A vision model selects the rendering strategy, a deterministic tool creates the image, and an independent quality gate grades and retries the result. Inputs, outputs, timings, and reasoning stream to a live dashboard.

DESIGN

Two independent LLM decisions separated from deterministic rendering and evaluation.

QUALITY

Reproducible image metrics trigger retries instead of relying on model confidence.

TRACE

Every pipeline event is recorded and visible down to the individual job.

  • NestJS
  • React
  • Ollama
  • BullMQ
  • sharp

02 / APPROACH

Evidence over
impressions.

01

Freeze the measure

Define the dataset, scoring, and thresholds before the run. If the target moves after the result, the evaluation is decoration.

02

Prefer deterministic signals

Use reproducible metrics where they are meaningful. An LLM judge is a tool—not a default answer.

03

Make systems observable

Store inputs, outputs, decisions, timings, and provenance. Improvement starts with seeing what actually happened.

04

Keep data local

Self-hosting is both an engineering choice and a privacy stance. The system should earn the right to move data elsewhere.

03 / ABOUT

Engineer first.
AI, measured.

I’m Ulad, an AI engineer based in the Netherlands. My background spans frontend engineering, technical leadership, and production AI systems.

Today I work on retrieval, structured extraction, agentic workflows, and the evaluation machinery that makes model changes measurable. Outside work, I run a dual-GPU home lab for local inference, vision, and speech workloads.

FOCUS AREAS

  • LLM evaluation & golden sets
  • Agent testing & quality gates
  • Local inference infrastructure
  • Mechanistic interpretability
  • Privacy-focused software

04 / CONTACT

Building something
that needs to actually work?

raer.camp@gmail.com ↗