Domain-specific AI is almost always bespoke: every plant, every line, every dataset gets its own model. PFN Studio is the workbench for the other approach — train one foundation model on synthetic priors that capture your domain, do in-context inference on anything it sees. TCPFN is one we built that way. This page is the case study.
Continuous-process manufacturing has the same question shape everywhere — what drives what, what happens if we change this setpoint, why did yield drop yesterday — but the sensors, lags, and causal wiring differ at every site. TCPFN is a single foundation model that handles that question shape across sites, trained on synthetic structural causal models with realistic process dynamics.
Sensor-to-sensor DAG from observational time series.
Per-instance CATE for any candidate intervention.
Counterfactual ranking of candidate root-causes on anomalies.
TCPFN's prior was authored in the studio's visual editor. Its architecture is a stack of transformer blocks from the studio's registry. Its training schedule used the studio's curriculum-scaling runner. Its evals are in the studio's eval suite. Its checkpoint is in the studio's marketplace. That's the full case study — the studio did the work that's usually six projects.
Recovers the canonical 11-node biological DAG (v2.1).
Held-out synthetic interventions with ground-truth effects (v3).
Simulated anomalies on AS/RS multishuttle case study.
Real physical testbed: 32 sensors, 30h walks, 28 RCT validations.
Numbers current to v2.1 / v3 training runs. The full eval methodology + ablations are in the white paper — book a call for an early read.
A model built in the studio doesn't have to leave it to ship. Checkpoints sit behind a versioned predict endpoint; client teams call it without redeploying. TCPFN powers Pearl, the digital-workforce demo from ProfitOps — the same team that built this studio. It's what a studio-trained model looks like once it's serving real industrial data.
Free during early access. Bring your domain knowledge; the studio brings the authoring, training, eval, and sharing primitives. Build a model that does in-context inference on data nobody else has seen.