- Client
- Global financial publisher
- Domain
- Publishing automation
- Method
- Deterministic NLG
700+ articles. Every day. Nobody writing them.
Building a lights-out publishing factory for global financial markets.
01
The problem
There are stories journalists should spend their time writing.
Routine market updates aren't necessarily among them.
A global financial publisher needed to produce large volumes of time-sensitive financial coverage across stocks, markets, and commodities.
Each story followed known editorial logic:
- What moved?
- By how much?
- Compared with when?
- Was it up or down?
- What historical context mattered?
The challenge wasn't creative writing.
It was producing accurate coverage for hundreds of instruments, across multiple markets and time zones, at exactly the right time every day.
At that scale, adding more writers wasn't a viable architecture.
02
The human loop
Without automation, every story would require some combination of:
find datainterpret movementdraftreviewpublishrepeat
Multiply that by hundreds of securities, multiple markets, several publishing windows, and multiple languages.
The volume wasn't merely expensive for humans to produce.
It was practically impossible.
03
What we built
Over roughly six months, we built an event-driven, serverless publishing system on AWS.
Live market data flowed into a deterministic natural-language-generation pipeline.
Instead of asking a language model to invent prose, the system evaluated market movements against predefined editorial rules and selected validated templates based on conditions such as:
- movement since the previous close
- intraday or historical changes
- positive or negative movement
- asset type
- market
- publication timing
- language
The resulting articles were automatically assembled and distributed to downstream publishing systems.
No writer.
No editor waiting in the loop.
No publish button.
Once the system was validated and running in production, it operated lights-out.
04
What changed
700+
articles/day
- generated and published automatically
Multiple
global markets
- covered around their local trading schedules
3
languages
- including English, Japanese, and Chinese
Near-zero
editorial error rate
- enabled by deterministic generation rather than probabilistic prose generation
Stories could publish pre-market, after market close, or around local-market schedules without an editorial team manually producing every update.
05
We automated the factory, not the newsroom.
The point wasn't to replace journalism.
It was to stop journalists from spending their mornings rewriting structured market data into predictable prose.
Human editorial expertise moved upstream.
We also built a custom templating platform that allowed non-technical editorial teams to author new article types and publishing logic themselves.
They could define the journalism.
The system handled repetition.
System → rules, templates, publish
Human → editorial logic, new story types
06
Under the hood
- AWS
- Serverless architecture
- Event-driven processing
- Live market data
- Deterministic NLG
- Rules engine
- Automated publishing
- Multilingual generation
- Custom templating platform
07
The lesson
Sometimes the right answer isn't more AI. It's deterministic software doing repetitive work extremely well.