SmartInferAbout SmartInfer
SmartInfer exists to make consequential decisions and actions measurably better.
Why specialist AI systems
High-value applications need more than plausible output from a general-purpose model. Generative models operate in an open-ended space; consequential systems must bound their outputs and actions by what the domain permits, what the evidence supports, and what can be independently checked.
The model is therefore only one component. In marketing measurement, statistical estimation, identifiability, uncertainty, and constrained optimisation must support the analysis. A shopping associate needs structured catalogue knowledge, retrieval, ranking, grounded answers, and controlled actions. Software engineering may require executable tests, acceptance gates, and formal verification where appropriate.
SmartInfer combines this domain-specific science with purpose-built agent infrastructure and a trust layer. The checking mechanism depends on the work: credibility checks for measurement, grounding and constraints for shopping, or tests, contracts, and formal methods for software. Verification does not mean applying formal proof everywhere; it means checking consequential outputs against criteria independent of the model’s own confidence.
Some specialists may also benefit from smaller specialised models when task-specific evaluation demonstrates advantages in cost, control, or performance.
The platform provides common foundations without forcing every specialist into the same architecture.
Further reading
The Missing Layer in Agentic AI → AI’s Edison Moment: The Moat Is the Grid, Not the Spark →
How we choose problems
We start with the economics of the problem, not the technology.
- Where is meaningful economic or operational value being lost?
- Which decision or action determines that outcome?
- Can a suitable AI system materially improve it?
- How will we know that it did?
We pursue a direction when there is a specific user, a plausible intervention, and a credible way to measure the result.
Why retail first
Retail provides a recurring scoreboard. Spend, revenue, promotions, discovery behaviour and customer response can be observed repeatedly, making it possible to test whether a system actually improves outcomes.
SmartInfer’s first commercial work is therefore in retail, beginning with marketing allocation through Galileo and extending to product discovery and product discoverability.
Founder
Anjan Goswami
Founder, SmartInfer
Anjan has spent two decades building and leading production search, ranking, machine-learning, and AI systems across Walmart, eBay, Adobe, Salesforce, Amazon, and Microsoft, most recently leading AI for PowerPoint Copilot.
He holds a PhD in computer science from UC Davis and an M.Tech from IIT Kanpur.
His technical interests span practical machine learning, programming languages, distributed systems, and mathematical representations of complex systems. He currently reviews for NeurIPS and SPLASH, and has previously reviewed for The Web Conference (WWW), KDD, ICDM, and Electronic Commerce Research and Applications. He writes occasional technical essays that approach technical questions from first principles.
SmartInfer grew from a recurring lesson across that work: a more sophisticated model is often not sufficient by itself. The decision, its measurement, and the surrounding system determine whether technical capability becomes useful.