Articles
Notes from our work
Short, practical write-ups on automation, integrations and AI — what worked, what did not and why.
-
What One Summarisation Pass Destroys
Summarise a corpus once, then answer from the summaries: one sentence in the summariser's prompt moved accuracy further than the choice of architecture, and compression discarded the names and thresholds before the decisions.
-
Retrieval, Full Context and Agentic Search on a Corpus of Superseding Decisions
1,527 logged model calls on a corpus of repeatedly amended decisions: the expected failure — confidently stale retrieval answers — never appeared, and where it finally did, it was full context that broke.
-
When does a blocking architecture verifier help a coding agent?
Three controlled studies, 732 agent runs, two model families: a blocking architecture gate helps only when its rules cover the bad solution shapes the model actually produces.
-
Cheap Infrastructure Does Not Mean a Simple Product
A small tool for watching new invoices in Poland's central e-invoicing system turned into lessons about edge runtimes, PDF generation and security promises you can honestly keep.