Why Senkovi?
I chose the name Senkovi after Disra Senkovi, a fictional terraforming scientist in Adrian Tchaikovsky's Children of Ruin.
What draws me to the wider series is how progress comes through research, experimentation, setbacks, and encounters with unfamiliar ways of thinking. Outcomes are earned through curiosity and practical work, not guaranteed by good intentions. That is what I want this site to document: the path from an idea to something built, tested, and improved.
About me
I'm Josh Angel, an AI researcher and engineer from Auckland, New Zealand. I lead AI research and model development at Straker Group. My work spans models, agents, software, and the infrastructure needed to run them.
I came to this work through DevOps and security. Servers, networks, deployments, incident response, hardening, and automation are still part of my technical practice. Building a system means understanding how it runs, not just how to get it started.
I usually start with a question or an idea. Research helps me work out what is already known and what is worth testing. Design turns that into something I can build and measure.
Running it brings different questions. What happens when a dependency fails, permissions drift, or the data changes? I investigate those limits, revise the design, and apply what I learn to the next iteration.
From automation to models and agents
My model work includes training runs, evaluation harnesses, data pipelines, and deployment. I want to understand what changes a model's behaviour and whether an improvement holds beyond a single example.
With agents, I work on memory, routing, tools, and feedback. The practical questions include what context to keep, what evidence supports a decision, and when a human needs to decide. Permissions and state matter as much as the model's answer.
In localisation, useful evaluation goes beyond fluency. Terminology, tone, and domain-specific language need to hold up across real workloads. That connects model research to the software and operational constraints around it.
At Straker I work on custom models and agentic workflows for professional translation and language operations. That includes small models tuned to customer domains, evaluation, and agents that route work and retain context.
Close to the metal
Outside work I train models from scratch on my own hardware. The current one is rivet-1b, a 1.09B parameter model with its own tokenizer, corpus, and post-training, trained locally on roughly 130B tokens. The weights and paper are on Hugging Face.
Training locally makes the trade-offs concrete. Data, compute, architecture, and measurement all affect what I can test. Failed runs and unexpected results help me decide what to investigate or change next.
DevOps and security remain central to this work. I build deployment paths, automate repetitive tasks, and investigate failures. Threat modelling and adversarial testing help me examine how a system can be misused as well as how it is meant to work.
I am interested in the whole path from an idea to a running system. Research informs the build. Operating it exposes assumptions. The next design should reflect what actually happened.
Why this site exists
Senkovi is my personal portfolio and publication. I write research notes, model reports, agent builds, software projects, and lessons from DevOps and infrastructure. Some pieces first appeared on the Straker labs blog; the rest is personal work.
I use it to document what I tried, what worked, what failed, and what I am still investigating. The aim is to make the reasoning and the implementation useful to someone working through similar questions.