SmartInferProducts
Current product · private pilots
Marketing allocation
Galileo — the Marketing Scientist agent
Decision: How should the next marketing budget be allocated?
Galileo is built around a stricter measurement question: not only whether a model fits historical outcomes, but whether its estimates preserve decision value. It models channel response and uncertainty, checks whether effects can be separated from the available data, and withholds reallocations the evidence cannot support.
An evidence-grounded AI agent lets teams ask the kinds of questions they would put to a statistician or economist—including model-based counterfactual questions about budget allocation—with answers bounded by computed results, uncertainty, and credibility checks.
Channel contributions · response and saturation curves · budget scenarios · confidence estimates · evidence traces
Research benchmark. Across 90 synthetic datasets with known economic ground truth, Galileo 0.2 recorded the lowest median normalised decision regret among five evaluated methods. In paired comparisons, it outperformed Google Meridian on 55 of 90 datasets, with two ties; Meta Robyn on 85 of 90; and PyMC-Marketing on 67 of 90. It did not lead every data regime.
Galileo 0.2 is a separate SmartInfer Research estimator. The benchmark supports Galileo’s decision-science direction; it does not establish real-customer lift or universal superiority.
Evaluation design. The benchmark used nine controlled data-generating regimes with ten deterministic seeds per regime. Because the true channel-response curves were known, each recommended allocation could be compared with the oracle allocation. The comparison held observations, train/holdout slices, budgets, constraints, and the allocation procedure constant across methods. Results include regime-level outcomes, bootstrap confidence intervals, and Holm-adjusted paired statistical tests.
Evidence & research
- From Model Fit to Marketing Decisions → Experimental report · 90 synthetic datasets · decision-quality benchmark
- The Confident Wrong Number → Research note · synthetic identifiability demonstration
Coming soon
Retail specialists
-
Shopping Associate AI agent Under development
Objective: How can a retailer help each customer accomplish their shopping objective through conversation and action?
A customer-facing conversation and action engine designed to behave like a human shopping associate within its permitted role. It talks naturally with shoppers, asks clarifying questions, searches the catalogue, and answers product, brand, and policy questions from grounded evidence. It can compare suitable options and use retailer tools to help process orders and returns, while stating when the available evidence cannot support an answer.
Underneath, SmartInfer’s agentic operating-system core manages memory and conversation state, invokes tools, grounds responses in catalogue and policy evidence, and governs permitted actions.
-
Product Discoverability Agent Under development
Objective: How should a catalogue or product listing be transformed so search engines, AI shopping systems, and LLM-based interfaces can understand and retrieve its products?
An action agent that accepts a catalogue or individual product listing and returns a cleaned, enriched version. It improves titles, attributes, taxonomy, descriptions, structured metadata, comparison facts, and product relationships so products can be interpreted and retrieved more effectively by AI discovery systems. Outputs can be formatted for straightforward upload to Google, Amazon, and Walmart.
Specialist models · private build
Domain expertise in smaller models
Question: When can a bounded domain be served by a smaller specialised model?
Domain Expert Private build
Technology for building smaller models with bounded domain expertise, evaluated against the tasks they are intended to perform.
Engineering · private build
Engineering decisions where verification matters
Generated software creates a second decision beyond generation: what evidence is sufficient to accept, merge, deploy, or certify the artifact?
Software acceptance
Decision: What evidence is sufficient to trust a generated software change?
Software Engineer Private build
An early direction for specification-driven generation, test-backed repair, and evidence-gated software changes.
Evidence & research
- Verified Software Generation: The First Protozoa Experiments → Experimental report · two early end-to-end systems · bounded verification and repair