The intelligence would be hybrid rather than model-monolithic. Vision-language systems resolve product identity across images, labels, receipts and changing descriptions. A commerce foundation model learns reusable representations of products, demand, inventory, attention, promotion and substitution across millions of sparse series, markets and modalities. It is not trained merely to predict next-week sales: it learns the latent grammar through which offers become comparable, missions form, demand moves and shocks propagate. Causal representation learning separates response to intervention from coincidental movement. Graph neural operators model substitution, complementarity and supply dependencies even when the graph changes. World-model agents maintain competing hypotheses about the market, generate interventions and update their beliefs when experiments disagree. Generative systems propose new formulations, claims, packs and portfolios, but neuro-symbolic constraint solvers reject ideas that violate regulation, unit economics, manufacturing or material limits. The system can use active learning to ask which new observation or micro-experiment would reduce decision uncertainty most. A calibrated adversarial critic continuously searches for distribution shift, impossible simulated behaviour, Goodhart effects and conclusions driven by missing data. Language is the interface to the model, never the source of truth: every explanation must compile back to evidence, equations, assumptions and observed experiments.
Living Shelf
“Can commerce become a living simulation rather than a sequence of guesses?”
The beginning
Living Shelf starts from a simple observation: an FMCG product is never sold in an abstract market. It exists simultaneously as a formulation, a pack, a price, a promise, a physical object moving through a supply network and a possible answer to a human need. Today those dimensions are optimised by different teams and different systems. I imagine something more ambitious: a continuously learning digital twin of commerce in which every product, store, channel, consumer mission, promotion, inventory position and external shock has an explicit place. Before changing a real price or launching a real pack, a company could create the decision inside this synthetic market, observe several plausible futures, understand who benefits or loses, and choose a bounded experiment. The aim is not to automate the extraction of value from consumers. It is to make the whole commercial system legible enough to design better offers, fairer pack architectures, more resilient supply and less waste.
The present market is observed through incompatible shadows. EPOS records what crossed a till, but not the intent that preceded it. Search and browsing reveal consideration but not household constraints. Shipment data describes flow into a retailer, not true consumption. Loyalty data is rich but partial and ethically sensitive. Market panels are carefully constructed but delayed. Digital shelves change by the hour; physical shelves are seen through intermittent audits; and availability is often inferred only after sales disappear. Meanwhile, pricing, assortment, innovation, media, trade promotion and supply planning run separate models of the same reality. Generative AI can produce hundreds of product concepts, yet there is no credible environment in which to ask which combination of formulation, claim, pack, price and route to market should exist. The deeper problem is epistemic: commerce has enormous data but no shared causal memory, no coherent representation of demand, and no safe place to rehearse decisions before imposing them on real people.
What it could become
Living Shelf would be a commercial world-model studio. Its first layer is a universal product graph: SKUs, ingredients, claims, sizes, bundles, substitutes, complements, stores, channels and supply constraints represented through time. Its second layer is a privacy-preserving demand observatory that combines transactions, availability, search, weather, events and consented consumer research without constructing dossiers on individuals. Its third layer is a population of synthetic market agents: not fake consumers, but calibrated behavioural hypotheses representing missions, budgets, habits, switching rules and uncertainty. Its fourth layer is a governed experiment engine. A team can generate a product or pack, place it into thousands of simulated market states, expose it to competitor and supply responses, identify failure regimes, and then design the smallest real-world test capable of resolving the most important uncertainty. The output is not a magic price. It is a decision envelope: feasible options, expected distributions of outcomes, causal assumptions, affordability effects, operational consequences, irreversible risks and an explicit learning plan.
For whom
- FMCG brands managing portfolio, revenue growth and trade investment
- Retailers balancing category value, availability and customer trust
- Category, key-account, e-commerce and demand-planning teams
- Smaller manufacturers needing rigorous market intelligence without a large data department
- Consumers who benefit from clearer packs, fairer unit economics and fewer stock-outs
Core capabilities
- SKU identity and product matching across retailers, languages and changing pack descriptions
- Comparable unit-price, pack, ingredient, claim and format intelligence
- Local assortment and distribution-gap detection by store or market cluster
- Baseline demand, price elasticity, substitution and cross-elasticity estimation
- Promotion incrementality, cannibalisation and post-promotion dip measurement
- Scenario planning across volume, revenue, margin, affordability, availability and waste
- Controlled commercial experiments with approvals, guardrails and learning capture
- A persistent commercial digital twin connecting product, demand, inventory, channel and context
- Synthetic populations for testing market hypotheses before exposing real consumers
- Generative pack, formulation and portfolio design constrained by cost, regulation and supply
- Privacy-preserving collaboration through federated learning, clean rooms and secure aggregation
- Real-time shelf sensing from e-commerce, store systems and computer vision
- Counterfactual stress tests for inflation, shortages, climate events and competitor response
The project becomes meaningful only when a new technical possibility is translated into a clear human advantage, an experience people can understand, and a system capable of earning trust over time.
Intelligence and mathematics
Mathematically, the market becomes a partially observed dynamic game played on coupled graphs: products and substitutions, missions and constraints, stores and catchments, factories and supply routes, and interventions through time. Structural causal models distinguish assumptions from relations identified by experiments. Hierarchical Bayesian models borrow strength while preserving local difference; state-space models reconstruct demand censored by availability; mean-field games approximate strategic response in large populations; optimal transport measures how demand migrates; and graph diffusion traces shocks. The more radical possibility is a differentiable market simulator: an approximate world model through which gradients can flow backwards from desired system outcomes toward candidate products, packs, prices, distribution and supply configurations. Instead of only asking what a proposed product will do, Living Shelf could ask which feasible product-system design is most likely to create a chosen region of outcomes. Because no simulator is reality, ensembles of structurally different models would compete, and conformal risk controls would define where the system is allowed to recommend an experiment. Distributionally robust optimisation searches for decisions that survive model error. A Pareto surface preserves incompatible objectives—margin, growth, affordability, retailer value, resilience, carbon and waste—so no political choice can disappear inside one score.
How it might live
The long-term company would be infrastructure for commercial simulation, not another pricing dashboard. The entry product could remain narrow—promotion incrementality and availability-aware elasticity—but each deployment would deepen the ontology, experimental language and market-response model. Brands could build portfolio futures; retailers could design categories in strictly separate environments; suppliers could model capacity shocks; and smaller companies could use a shared foundation without surrendering transactions. Federated learning, confidential computing, differential privacy and clean-room computation could improve collective representations without exchanging raw data or future commercial intentions. A cryptographically signed decision object could carry the model version, authorised evidence, assumptions, constraints, human approvals and evaluation plan into execution, producing an auditable lineage from hypothesis to shelf outcome. Over time, Living Shelf could become a programmable institutional layer: markets specify machine-readable rules for affordability, price stability, promotion frequency, carbon, essential availability and competitive separation, then verify that every proposed action remains inside them. The moat would not be one algorithm; it would be the interaction between ontology, causal memory, calibrated simulation, governance protocol and execution feedback.
For me, a venture is more than an interesting technology. It needs a narrow first user, a repeated problem, a distribution path, a credible advantage and a reason to improve as more people use it. I would test those conditions before deciding whether this idea should become a company, a product, an open technology or an ongoing research programme.
Rules for making it real
- 01
Optimise the offer, not the consumer's vulnerability.
- 02
Treat pack, price, promotion, placement and availability as one system.
- 03
Show unit economics and consumer unit price together.
- 04
Never confuse a stock-out with a lack of demand.
- 05
Prefer bounded experiments to confident universal recommendations.
- 06
Keep customer data isolated and prohibit competitor coordination or future-intention exchange.
- 07
A person approves every consequential price or assortment change.
- 08
Simulated people are hypotheses to test, never digital replicas to exploit.
- 09
Every recommendation must travel with its assumptions, uncertainty and expiry date.
- 10
The model must preserve disagreement between commercial, consumer and environmental objectives.
- 11
A world model earns authority only through prospective experiments and explicit failure boundaries.
- 12
Rules governing commerce should be machine-readable, inspectable and cryptographically attributable.
- 13
Optimise systems and offers; never optimise an identifiable person's moment of weakness.
From question to company
- 01Frame
Choose one category and market, build the product identity graph and establish trustworthy unit-price histories.
- 02Prototype
Connect sell-in, sell-out, distribution, stock and promotion data for a small group of design partners.
- 03Prove
Measure baseline demand and promotion incrementality, comparing recommendations with existing planning methods.
- 04Build
Run controlled assortment and promotion experiments with pre-agreed consumer, margin and availability guardrails.
- 05Advance
Add price and pack architecture scenarios only after the causal measurement layer is credible.
- 06Advance
Build a calibrated synthetic market and prove that it predicts experiment direction—not just historical fit.
- 07Advance
Introduce generative pack and portfolio design only after the simulation can reject its own attractive mistakes.
- 08Advance
Add federated market learning without exposing raw participant data or competitive intentions.
- 09Advance
Train a multimodal commerce foundation model and test whether its representations transfer to unseen categories and shocks.
- 10Advance
Prototype a differentiable simulator that reasons backwards from outcomes to feasible pack-and-market designs.
- 11Advance
Create signed decision objects and a policy engine that can prove which data, model, rule and human approval produced an action.
- 12Advance
Expand across categories and geographies while maintaining separate models, legal review and local commercial logic.
What could go wrong
Serious imagination includes the possibility that an idea should change radically—or should not exist. These are the tensions the project would need to resolve:
- Bad product matching creating false price comparisons.
- Historical data encoding promotional habits that should not be repeated.
- A price recommendation being mistaken for certainty rather than a testable estimate.
- Optimisation improving margin while quietly reducing affordability or pack fairness.
- Competitively sensitive data creating antitrust or price-coordination risk.
- Retailer and manufacturer objectives being collapsed into a single misleading metric.
- Local recommendations becoming discriminatory proxies for individual willingness to pay.
- Teams automating execution before governance and causal evidence are mature.
- Synthetic consumers reproducing stereotypes while appearing scientifically precise.
- A digital twin drifting from reality yet retaining institutional authority.
- Generative product abundance creating complexity, waste and false novelty.
- Simulation becoming a substitute for listening to real consumers and frontline teams.
- A foundation model concentrating market knowledge and power in one private infrastructure provider.
- Differentiable optimisation finding commercially effective but socially perverse shortcuts.
- Privacy technology being used as a legal-looking wrapper around excessive data collection.
- Machine-readable policy creating false confidence that contested social values have become technical facts.