Ship
Containers, CI, observability. The unglamorous half that decides if it is real.
Three projects
How I work
A demo proves an idea can work once. Shipping is the argument that it will keep working when you are not in the room, on inputs you did not choose, for people who will not file a bug report.
For NeuroClima that meant a Kubernetes cluster underneath: networking, storage, vLLM serving, autoscaling, and a pipeline that wakes up when a document lands in S3. Getting a response from sixty seconds down to ten took longer than building the first working version, and improved the answers as a side effect.
The part people skip is instrumentation. If you cannot see what real queries look like you are guessing about what to fix, and the thing you guess is almost never the thing that is wrong.
NeuroClima
The cluster underneath a live EU platform: vLLM serving, ingress, database config, autoscaling, an S3-triggered ingestion pipeline, and response tracing so administrators can see real sessions rather than a synthetic dashboard.
View project ↗AIKAA
Roles, activity logs and fallbacks go in from the first demo, and it still gets called a demo until the company has used it. Real use turns up the edge cases, and those get worked through with them.
View project ↗Voice to Value
An app on fourteen people’s phones for their real lunches, which is where you find out that the failure modes you planned for are not the ones that happen.
View project ↗