Insights

Notes from the build.

Writeups from work we've actually done: the decisions, the things that broke, and the numbers that came out. No trend pieces.

3 notes

14 July 20266 min read
Most RAG systems fail at retrieval, not generation
When a retrieval-augmented system gives a bad answer, the instinct is to change the prompt or the model. Usually the right passage was never in the context.
- RAG
- LLM systems
- Evaluation
9 June 20267 min read
Why we split ML inference into three services
Ingest, inference, and delivery fail differently and scale differently. Putting them in one process means the slowest one sets the rules.
- System design
- Kubernetes
- Kafka
18 May 20266 min read
Cutting a 600 MB model app down to 70 MB
ResoNet needed to run respiratory screening entirely on-device, in places with no reliable network. The model was never the problem — the runtime was.
- Edge AI
- ONNX
- Mobile

Let's talk

Have a problem worth solving?

Tell us what you're building. We'll help you scope it, build it, and ship it.

Tell us what you're building

or email hello@somokolonlabs.com