W23: Autonomous News Engine, LV Quality System, and Model Cost Optimization
A week where three systems merged into one story about reliable automation — an autonomous daily news engine, a multi-layer Latvian quality system, and continuous model cost optimization.
- Autonomous daily news engine — from source to publication with no human involvement
- Multi-layer Latvian quality system — a sieve bad text can't pass through
- Continuous model cost optimization — measure, don't guess
- Automation is only as good as the rules behind it
- A quality check has to be a gate, not a suggestion — otherwise it gets ignored
- Optimization is moving, not fixed — a choice found once goes stale
- Automatic publishing without quality gates is too risky — one bad day would reach the public
- A vague input standard gives a vague result — clarity in the rules is a precondition, not an afterthought
Week overview
Three systems, one story — how to make automation reliable. The news engine shows speed, the Latvian quality system shows the safety net, and cost optimization shows discipline. Each is useful on its own, but together they form a principle: routine can be handed to a machine, as long as the human keeps the standard and the check.
Autonomous daily news engine
We built a pipeline that runs the whole path on its own every morning — gathering industry sources, filtering noise, drafting business news in Latvian, and publishing it. The human’s role here is to set the editorial standard, not to sit down at the keyboard each morning.
The biggest gain isn’t speed but consistency. The system never forgets to publish and doesn’t hold one standard on Mondays and another on Thursdays. The human time that used to go to routine is freed for strategy.
Multi-layer Latvian quality system
Language models write fluently, but in Latvian fluency isn’t the same as naturalness. So we built a system with layers rather than a single check: a forbidden-word check, morphology, and a style rating done by the model itself. Together they form a sieve the feel of machine translation can’t pass through.
This system became the foundation for others that write in Latvian. Without such a safety net, automatic publishing would be too risky.
Model cost optimization
The third system measures, it doesn’t guess. For each type of task we compare several models on two metrics at once — how good the result is and how much it costs. The goal isn’t to find the best model in the abstract, but a good-enough one at the lowest price for each job.
Continuity matters. The model market moves fast, so a choice found once goes stale quickly. A system that re-evaluates regularly keeps its edge.
Mistakes and lessons
No gates, no quality. As long as a check only recommends, it gets ignored. Only once it becomes a condition for publication does quality actually hold its level. Automatic publishing without such gates is too risky.
Vague input, vague result. When the editorial standard was weak, the system reflected it precisely. Clarity in the rules turned out to be the most important investment — a precondition, not an afterthought.