Brain ideas

This page exists because the most common question after “what’s a prior?” is “what could I actually build with this?” Below is a working menu of brains, organised by role. None of these are products — they’re shape-of-problem sketches you can take into the wizard, swap your own features for, and train.

The pattern across all of them is the same:

• The prior captures the structural regularities of the problem (e.g. “matching invoices have similar amounts and dates”).
• At inference time you pass a handful of your org’s labelled examples as context (e.g. 50 matches your AP team approved, 50 they kicked back).
• The same brain weights serve every team — only the context differs.

That’s the PFN payoff: train once, adapt without retraining.

Is my problem PFN-shaped?

A problem is a good fit if all five are true:

1. It’s a high-volume recurring decision — thousands per year, not one-off
2. The features are structured — extractable from an ERP / CRM / data warehouse, not raw text
3. There’s some labelled history — past decisions you can use as context
4. Org policy or context drifts — what counts as “OK” depends on the company / team / quarter
5. The decision is reversible or auditable — a human can override a wrong call

If those don’t hold, a PFN brain is probably the wrong tool. Free-form text reasoning, one-off strategic calls, hard real-time control loops, or problems with millions of labelled rows already (use XGBoost) all fall outside the sweet spot.

For a CFO / Finance org

Accounts payable, close, treasury, FP&A — the highest-density use cases on the catalogue today.

Brain	Job-to-be-done	Capability
Invoice matching	Match invoice ↔ PO ↔ receipt; flag variances by reason (price / quantity / vendor typo / off-cycle)	Sort into categories ⏳
Duplicate-invoice detector	Same vendor, similar amount, close date — already paid? Catches bot-resubmitted bills + manual double-entry.	Sort into categories ⏳
JE anomaly screen	Flag journal entries unusual vs this org’s history; runs nightly during close so the controller sees the list at 9am	Sort into categories ⏳
Expense → GL classifier	Auto-tag expense submissions to GL / cost-centre; learns from corrections	Sort into categories ⏳
Approval routing	Pick approver tier from amount + category + vendor history + budget remaining	Sort into categories ⏳
Cash-flow forecast	Daily/weekly cash position from historical inflows/outflows + scheduled flows	Forecast forward
Collections aging	When will this open AR get paid? + how confident are we?	Forecast forward + Show confidence
Customer credit risk	Default likelihood from receivable + ordering signals + payment history	Sort into categories ⏳
Vendor health screen	Supplier-side financial-stress flag from payment + delivery patterns	Sort into categories ⏳
Budget-vs-actual ETA	Where will this cost centre land at month-end given current run-rate?	Forecast forward

Where to start. JE anomaly screening or invoice matching. Both are high-frequency, both have labelled history (every JE was eventually accepted/rejected; every invoice was eventually matched or kicked back), both touch the close process where ROI is most visible.

For an Operations / Plant Manager

Industry 4.0 — sensor data, production lines, equipment uptime.

Brain	Job-to-be-done	Capability
Throughput forecast	Predict next-shift output from current line state + historical patterns	Forecast forward
Quality drift detector	Flag SPC excursions before they breach control limits	Sort into categories ⏳
Predictive maintenance	Time-to-failure estimate from vibration / temperature / runtime signals	Forecast forward + Show confidence
Setup-time predictor	Given changeover details, predict the realistic line restart time	Predict a number
Bottleneck attribution	Which station drove last shift’s miss? (requires discovery — not in catalogue today)	—
Shift-handoff anomaly screen	Flag readings the outgoing shift normalised but shouldn’t have	Sort into categories ⏳

For a Sales / RevOps org

Pipeline, forecasting, account-level attribution.

Brain	Job-to-be-done	Capability
Deal-stage forecaster	Probability this opp closes this quarter, given activity signals	Sort into categories ⏳ + Show confidence
Pipeline coverage forecast	Quarter-end ARR given current pipeline + win-rate history	Forecast forward
Lead-routing classifier	Route inbound to best-fit AE based on territory + segment + history	Sort into categories ⏳
Churn early-warning	Flag accounts whose engagement pattern matches past churners	Sort into categories ⏳
NRR forecast per cohort	Expansion + churn forecast for a customer segment	Forecast forward

For an HR / People org

Recruiting funnel, attrition, comp.

Brain	Job-to-be-done	Capability
Time-to-fill estimator	Days-to-hire forecast from req attributes + funnel state	Predict a number + Show confidence
Offer-acceptance scorer	Likely-to-accept given offer terms + candidate signals	Sort into categories ⏳
Attrition early-warning	Flag employees matching past-leaver patterns	Sort into categories ⏳
Comp-band fairness audit	Spot offers / promotions far from internal cohort	Sort into categories ⏳

For an Engineering / ML org

Meta-uses — using PFN Studio on the ML workflow itself.

Brain	Job-to-be-done	Capability
HPO surrogate (BO)	Pick next hyperparameter combo from results so far	Pick the next experiment
Learning-curve extrapolator	Cancel hopeless training runs early	Extrapolate a partial curve
Freeze-thaw BO	Multi-fidelity HPO across many candidate configs	Pick the next experiment + Extrapolate
Incident-cause classifier	Tag postmortem root causes from telemetry windows	Sort into categories ⏳
Cost-of-experiment forecast	Predict the GPU bill before kicking off a run	Predict a number

Build path

Pick a brain idea, then:

1. Sketch the features. A list of structured columns that come out of your source-of-truth system (ERP, CRM, sensor table). Pin a feature count D and a row count N per task.
2. Find the closest catalogue capability. Use the capabilities reference — every brain above maps onto one of the six.
3. Open the wizard at /teach, pick the capability, get a starter brain.
4. Edit Prior on the brain page to swap in a synthetic generator that matches your feature distributions (the embedded Python in the prior file is the unique work; the model + eval stay the same).
5. Train on Vast.ai GPU for production brains (5-60 min); local CPU is fine for prototypes.
6. Pass your org’s labelled context at predict time. The /predict endpoint takes a { context: { x, y }, query: { x } } payload — the brain reads org-specific policy from the context every call.

The shorter version of all of the above: the brain is reusable; the prior captures structure; your data is the context.

What PFN Studio can’t help with (yet)

Honest list of where these ideas would need work the catalogue doesn’t ship today:

• Causal attribution / “why?” questions — discovery heads exist as a block type but no end-to-end paper-backed template for tabular CFO/ops causal discovery yet.
• Unstructured text reasoning — contract review, analyst notes, free-form ticket triage. PFN Studio is for structured signals after extraction.
• Sub-millisecond inference — a PFN forward pass is fast, but heavier than a linear model.
• Problems where you already have millions of labelled rows — XGBoost / your existing pipeline likely wins.

These limits will shift; track the roadmap for what’s coming.