.png&w=828&q=75)
Thinking Machines is a marketing intelligence platform built on a contrarian thesis: in 2026, every AI marketing tool is a fragile black box that gives different answers to the same question. Enterprise marketing teams can't deploy them in regulated environments. We took the inverse approach — LLMs as compilers, statistics as the workhorse. When a marketer asks "why did our DACH CAC spike in October?", a two-stage compiler (Gemini Flash for classification, escalating to Pro only for novel queries) translates the question into a structured plan of deterministic tool calls. Those tools are mostly battle-tested classical ML: Prophet for forecasting, Markov chains for attribution, Isolation Forest for anomalies, Bayesian structural time-series for causal inference. A governed semantic metric layer ensures "CAC" resolves to the same versioned, owned definition every time — never inferred by the LLM. A find_historical_analog tool grounds every diagnosis in past precedent ("this looks most like October 2024, when CAC normalized after a LinkedIn audience refresh"). When a question is genuinely novel and no existing tool applies, Investigation Mode activates: a SOTA model writes a new deterministic tool with type signatures and test cases, the human approves it, and from that point forward the new case is also deterministic. The system's kernel grows; trust compounds. Every answer carries a complete audit trail — tier classification, metric resolution, tool calls with parameters. Same question, same answer, always. The platform plugs into BigQuery, Confluence, HubSpot, Salesforce, and Slack out of the box, making it deployable in any mid-sized B2B SaaS company in an afternoon.
19 May 2026