The hard parts,
written down.

A founder writes one sentence; the system must produce an ordered, reversible chain of actions that respects a margin floor and a budget ceiling. Most planning research optimises for task success. We optimise for not breaking a rule the user set — a plan that earns more but violates the margin floor is a failed plan.

• Constraint satisfaction over long horizons
• Provable rollback for every committed action
• Cost-aware search under a spend ceiling

Scoring 200 markets continuously means deciding what to ignore. Demand, competition, acquisition cost, tax regime, language and seasonality go in; brand sentiment and social virality deliberately do not — they proved unstable across quarters and inflated scores in markets we later had to close.

• Signal selection and stability across quarters
• Transparent 0–100 scoring, reviewable by the founder
• Counterfactual evaluation of markets not opened

Invoicing, VAT, legal notices and GDPR are usually bolted on after the product works. We treat them as constraints inside the planner, so an action that would produce a non-compliant state is never proposed in the first place.

• Jurisdiction-aware action filtering
• Automated generation of legal surfaces per market
• Continuous verification rather than launch-time checks

The gap between intent and action changes what a founder is willing to try. Below roughly half a second, people experiment; above two seconds, they batch requests and stop exploring. We treat 0.4 s intent-to-action as a hard target, not a nice-to-have.

• Speculative execution of reversible steps
• Streaming plans before they are complete
• Cost of latency measured in experiments not run

Working on any of these? We collaborate with labs and hire from them — see open roles or write to research@getnallo.ai.