What we're building, what we're seeing
Graph engineering for agentic workflows
There's a graph inside your agent loop whether you drew one or not. Draw it, and the next time one hangs you'll know which node it stopped on.
Read moreHow Paradox Machines maps data architecture
Most companies can't answer who owns their data, what it's for, or whether it can be trusted. A four-layer reference architecture is the map we use with customers to help answer those questions, or identify gaps.
Read moreThe harness matters more than the model
Teams point a chat model at SQL, watch it fail, and conclude AI can't do data modeling. The evidence says they measured the wrong thing: the harness matters more than the model.
Read moreMatch the scorer to the job
Most eval suite failures aren't the agent — they're an LLM judge second-guessing a verdict a deterministic check already settled. Three scorers, each kept in its lane.
Read moreThe Prompt Isn't the Product
Prompts and tools are where you start. Tracing, eval, and iteration are where the improvement actually compounds.
Read more
