Judge whether AI output is actually good
Not only whether the facts are right, but what a reader would conclude from it. I caught a chart in my own material that was completely accurate and taught the opposite of its lesson.
AI & Technology Specialist · Pretoria, South Africa
Building Practical Technology Solutions at the Intersection of AI, Automation and Innovation.
Independent technology builder exploring generative AI, intelligent workflows, digital products and emerging software technologies.
Open to — remote, available immediately
- AI Quality Rater
- AI Evaluator
- AI Trainer
- AI Content Analyst
- AI Operations Specialist
- IT Support Specialist
- Software Testing Assistant
- Junior AI Engineer
- 2 desktop applications released Built on locally hosted AI models
- 20 episodes produced in 4K Unattended, from structured data
- 38 automated tests on one engine Dependency-free, one command
- 6 times I discarded my own results Documented, with the evidence
What I can do
Four things I would bring on day one
Each one links to the evidence behind it, including the parts that did not work.
Design an evaluation that survives scrutiny
Held-out splits with an adoption rule, instrument checks before hypothesis tests, and fixed randomness so that two runs are comparable at all.
Build and run AI systems end to end
Speech recognition, synthesis and language models running locally, wired into pipelines that run unattended and refuse to deliver output failing a check.
Keep a working system alive
A GPU workstation under sustained load, long unattended jobs, root-cause diagnosis, and a formal IT support qualification behind it.
Selected work
Six systems, built and shipped
Each case study covers the problem, the architecture, what I was responsible for, what went wrong, how it was tested, and what is not finished.
Operational
AI Product Development
Hoshi Nova
AI-assisted animation production platform
Twenty episodes across two seasons were produced end-to-end with this pipeline and published.
- 20Episodes produced
- 3840×2160Output resolution
- 11:02Longest episode verified
Released · early version
AI Product Development
Hoshidub
Live AI dubbing for anime, running locally
A released Windows application built entirely on locally hosted speech and language models, with no network dependency at runtime.
- 0.1.0Released version
- 3Models running locally
- 100% on-deviceProcessing location
Built · not yet proven live
AI Systems & Agents
Project Luna
Autonomous AI presenter and live-stream host
Built and functional in testing. It has never run a real public broadcast, and this page says so throughout.
- 0Operator actions required
- LLM + TTS + ASRModels running locally
- 2Platforms integrated
Delivered
AI Systems & Agents
FRIEND//DEBT simulation engine
Deterministic social-reasoning engine with an enforced information boundary
The engine was built as a game, validated by 38 tests, and ended up as the story generator for a published ten-episode season.
- 38Automated tests
- 0Runtime dependencies
- 8Simulated characters
Built · not yet proven live
Web Engineering & Tooling
This portfolio
Zero-dependency static site generator and publishing toolchain
The one project here you can inspect completely — the source, the build, the gate and the output are all in front of you.
- 0Runtime dependencies
- 18Pages generated
- ~50 msBuild time
Built · not yet proven live
AI Product Development
Hoshi Cut
Video analysis, scoring and short-form editing from one upload
An editing assistant whose entire design rule is that every judgement must trace back to something measured in the footage.
- ~100 sAnalysis speed
- 6Scoring scales
- 6Edit modes
Working method
I build software with AI assistants, and I verify every result before I claim it.
I am not a self-taught programmer who writes every line unaided, and this portfolio does not pretend otherwise. What I do is specify, direct, test and validate systems - then reject my own work when the evidence does not support it. The examples on the Verification page are all cases where I found the fault in my own results.
Education
Qualifications
-
Higher Certificate in Information Technology Support Services
IIE Rosebank College
Completed March 2024
-
INF4831 – Information Security
University of South Africa (UNISA)
NQF Level 8 · Distinction · 2021
-
Google Analytics Individual Qualification
Google
Certified · Certified
Currently looking for remote work
AI evaluation and quality, AI operations, technical support, software testing, or a junior AI engineering role. The career page sets out where I fit and where I do not.