// field notes

Know where it breaks.

I make large backend systems boringly reliable, and I test the claims around AI agents, local models, and self-hosted infrastructure until I find where they give. These are the write-ups.

The Philosophy

./where_it_breaks

The best infrastructure work is mostly invisible: teams ship faster because production stopped surprising them, and big changes land in live systems without anyone noticing the landing.

Now AI has made generating code and configs cheap, while understanding what they do to a live system is as expensive as ever. So I trust controls that enforce over instructions that ask: a key or a role binding is a gate that holds whether anyone is watching; a paragraph in a system prompt is only a weight. I'm glad to hand a machine the typing. I don't hand it the understanding.

./core_focus

dns
dns

Boringly Reliable Systems

Large-scale data transformation pipelines and the backend services around them. Shipping multi-quarter migrations and architectural changes into running production systems without waking anyone up.

bolt
bolt

Velocity Without Breakage

Cutting operational overhead and maintenance toil so teams ship faster without putting production at risk. Friction is worth removing, except where it turns out to be load-bearing.

science
science

Finding the Limits

Testing claims on real hardware and live clusters. I push agents, local models, and self-hosted stacks past their advertised promises, then write up what held, what failed, and what the numbers actually said.

./experience

Google

Staff Software Engineer
Nov 2019 – Present · Los Angeles

Technical lead for large-scale data transformation pipelines and supporting backend services that power advertiser traffic forecasting and planning tools. Evolving production pipelines in place, landing multi-quarter migrations without disruption, and cutting operational toil so teams ship on solid foundations.

Amazon

Software Development Engineer
Apr 2017 – Nov 2019 · Orange County

Full-stack owner of grocery recommendations and substitutions: high-throughput Java and Python services, ML models productionized with data scientists, and web and React Native interfaces, kept fast and stable through peak demand.

./selected_writings

Each piece starts with a question I couldn't leave alone and ends wherever the numbers did.

Personal writing. Views are my own and do not represent my employer.

01 ./local_ai_and_limits · 4 essays

Local AI, Hardware & Precision

Testing local weights, scrap hardware, and custom multi-GPU builds against the claims to see what the numbers actually say.

Bits of π
Quantization
Nov 2026 14 min read

Bits of π

The Short Version

Ollama drops three-quarters of a model's bits by default. Squeezing an image that hard yields garbage, yet 4-bit quants retain 95% accuracy. A 64-build experiment showing how block scales and mixed precision work—and why the loss lands on tokens the model was never sure about, starting at digit 813 of π.

02 ./agent_experiments · 2 essays

Agent Experiments & Comfort Boundaries

Experiments in pushing the boundaries of my own comfort: letting an agent into my homelab cluster and figuring out by trial and error what I'm actually willing to let it touch.

03 ./on_the_industry · 3 essays

Thoughts on the Industry

Where software titles, hiring rituals, and small software businesses land when the typing is automated.

How Hard Could the Magic Be
Build Log
Jun 2026 12 min read

How Hard Could the Magic Be

The Short Version

A weekend, $73 of API calls, and a working clone of a sketch-to-CNC product. The threat isn't the hobbyist who stops once the itch is scratched. It's what happens to any business whose main defense was “this was hard to build.”

The Punch Card Test
Hiring
May 2026 8 min read

The Punch Card Test

The Short Version

The cost of producing code has collapsed. The cost of understanding it has not. Interviews still measure the first one, and AI turned an old noisy filter into a compromised one.

Interactive Experiments & Open-Source Stacks

./lab_and_tools

Playable visualizers, benchmark harnesses, and homelab software built while testing the claims.

Aries Image Studio
palette Self-Hosted Image Studio
Live Web App

Aries Image Studio

Self-hosted studio for batch image generation, workflow orchestration, and multi-GPU resource management across local hardware.

Bits of π Benchmark Explorer
monitoring Dataset & Lab Report
Interactive Report

Bits of π Benchmark Explorer

Full experimental report and raw telemetry across 64 GGUF builds of Qwen (4B, 9B, 27B), tracking exact digit recall and teacher-forced KL divergence down to 2 bits.

Quantization Lightbox
tune Interactive Visualizer
Live Web App

Quantization Lightbox

Treat an image's color channels as a weight tensor. Inspect how GGUF Q4_0, Q4_K_M, block scales, mixed precision, and error diffusion behave at the pixel level.

Aries Dual-B60 Inference Stack
memory GPU Compute Stack
Open Source

Aries Dual-B60 Inference Stack

Dual-backend llama.cpp (Intel oneAPI SYCL + Mesa Vulkan) and vLLM tensor-parallel Docker stacks tuned for 2× Intel Arc Pro B60 (48 GB VRAM) on consumer PCIe boards.

Orbit Container Registry UI
deployed_code Container Registry UI
Open Source

Orbit Container Registry UI

Lightweight read-only web UI for private OCI and Docker Distribution V2 registries in my homelab, inspecting repositories, multi-arch manifests, and layer histories.

Kiryl Dubarenka

./the_author

Kiryl Dubarenka

Staff Software Engineer

A software engineer working at the seam between human judgment and machine-generated code. By day, I lead backend infrastructure at Google, keeping large-scale distributed systems dependable and quiet. In my free time, I run a homelab cluster with an AI agent in it, test the limits of self-hosted and agentic tools, and write about what I find.