Consulting

Behind Kompismoln is a developer who also takes on assignments. What I do best are systems where language models, data and operations meet, and that have to hold up when someone else examines them.

AI pipelines

Language models in your own workflows, from data in to answers out. The models can run locally with vLLM or be called behind OpenAI-compatible APIs, chosen by what the data can bear. Kompismoln runs open models on its own GPUs, so that sensitive data doesn't have to leave the organization.

Research tools: ChatDDX

ChatDDX is a rig that measures how accurately language models produce differential diagnoses and management plans from emergency-care case vignettes.

Language models aren't deterministic, and a model behind an API can change overnight. So everything that can change an answer or its score – prompts, sampling, model weights, cases, seeds and scorers – is pinned by its digest, so that a run can be rebuilt and examined by others. Answers are scored with pattern matching and with language models as judges, and sensitive vignettes are handled as fingerprints, resolved only where that's cleared.

ERP integrations

Connections between business systems and the rest of the organization: import and export, synchronization and reports, built so that you can see what happened when something goes wrong.

Infrastructure as code

Kompismoln runs on o11n, a framework where a single file describes a whole organization: servers, accounts, mail, networks, backups and workstations. Everything is built with Nix and can be recreated from scratch. The same way of working can be set up for others.

Contact

Write to info@kompismoln.se and tell me what you need.