Retail & Commerce
Catalogue quality is the ceiling on every model you ship
Search, ranking, forecasting and service in commerce all inherit the quality of one asset — the catalogue. Digiaeon builds the enrichment, retrieval and decisioning layers underneath, sized for a festive-sale peak and priced against contribution margin.
Retail & Commerce
Catalogue, search and demand systems that hold their budgets on the biggest sale day of the year
- Now
- 3 applications
- Next
- 3 applications
- Frontier
- 1 application
- Constraints we design around
- 5
Context
Three forces reshaping this sector.
What changed recently enough that the old architecture is now the wrong one.
Discovery is leaving the search box
A growing share of product research now starts inside an assistant rather than on a storefront. That surface reads structured feeds, schema.org Product markup and public APIs — not carefully tuned merchandising rules. Retailers that treat the product record as an internal artefact become invisible to it, and the usual session-based attribution stops explaining where demand came from.
Quick commerce reset the latency expectation
Ten-minute delivery pushed the planning grain from region-week to dark-store-hour, and made available-to-promise a real-time question rather than a nightly batch. Once a shopper has been served in ten minutes, a two-day promise on the same category reads as a defect. Every forecasting and inventory system in the business is now judged against that bar.
Inference became a COGS line
The growth-at-any-cost era paid for experiments out of the funding round. In 2026 an LLM call on the search path is measured against contribution margin per order, and at a few crore queries a month even a cheap model is a visible number on the P&L. The interesting engineering is no longer whether a model helps — it is whether it helps at a cost the category can carry.
Applications
What we would build here, in order.
Grouped by how close each one is to production today — not by how impressive it sounds in a board pack.
Build it this quarter
Current models, current data. The engineering is understood and the risk is schedule risk, not research risk.
Catalogue enrichment and attribute extraction
A vision-plus-text pipeline reads supplier images, spec sheets and seller free text, and writes normalised attributes against a real taxonomy — GS1 GPC or the retailer’s own. Extraction is confidence-gated: high-confidence values write straight through, the rest queue for human review, and every value carries its source so a wrong colour can be traced back to the PDF it came from. The same attribute record becomes the single grounding source for generated PDP copy, so marketing text cannot claim a specification the catalogue does not hold.
Hybrid retrieval, query understanding and ranking
BM25 and dense vectors run in parallel and fuse by reciprocal rank fusion, then a cross-encoder reranks a shortlist — fifty candidates, not five hundred. Query understanding handles the parts a vector model quietly fails at: spelling, brand disambiguation, and constraint extraction from things like “red running shoes under 3000”, which becomes a filter rather than a semantic hope. Null-result recovery matters more than average relevance; a zero-result query is a lost session, not a low nDCG.
Order-aware service deflection
Most contact volume in commerce is where-is-my-order, returns eligibility and policy questions. A grounded answer engine with tool access to order state, the returns policy of record and the carrier scan can resolve those without a human, provided it refuses cleanly when the tools disagree. The design rule is that the assistant never asserts a delivery date it cannot cite, and escalation carries the full context so the agent does not restart the conversation.
One thing has to change first
Buildable, but gated on something outside the model — labels, a consented holdout, an integration, or a process the business has to agree to change.
Demand forecasting and replenishment
Forecasting at store-SKU-day is mostly an intermittent-demand problem, not a time-series showcase. Long-tail SKUs sell in ones and zeroes, so Croston-style and count-based methods beat a deep model that has nothing to learn from. Festival calendars, promo flags and competitor stock-outs enter as regressors, hierarchies are reconciled so store forecasts sum to the DC plan, and the output is a service-level decision under a newsvendor objective — not a point forecast handed to a planner to argue with.
Price, promotion and markdown decisioning
Elasticity estimated from observational price changes is usually estimating the merchandiser’s intuition, because prices moved for reasons correlated with demand. Honest work here means structured experimentation, geo or store holdouts, and causal estimators — then a policy that lives inside hard guardrails: price bands, MAP obligations, and a floor that respects landed cost. Markdown optimisation over a season, with sell-through and residual inventory as the objective, is where the money actually is.
Returns and promotion abuse detection
Serial returners, coupon farming and wardrobing are graph problems before they are model problems. Entity resolution across accounts, devices, payment instruments and addresses surfaces the ring; the model scores it. The hard part is the action: a false positive here blocks a good customer at checkout, so the system produces a case with evidence for adjudication rather than a silent block, and the threshold is a business decision that gets reviewed.
Scope it as research, with a gate
Credible and worth a prototype. We would run it behind an evaluation gate and say plainly if it does not clear the bar.
Agent-readable commerce and the merchandising desk
Two things arrive together. Machine buyers — an assistant shopping on a customer’s behalf — need a catalogue that answers structured questions about availability, variants and returns, and a bot-management posture that can tell a customer’s agent from a scraper. On the supply side, an operations desk of agents drafts the replenishment exception, the markdown proposal and the supplier chase, each as a reviewable action with a rollback. The merchandiser stops typing and starts approving.
Constraints
What makes this sector different.
The part of the build that does not transfer from anywhere else. Every one of these has to be answered in the architecture, not in the project plan.
Peak is not a bigger Tuesday
Festive sale and Black Friday traffic arrives in minutes, at multiples an autoscaler cannot warm into. The failure modes are specific: cache stampede on a hot SKU, connection pool exhaustion at the database, a reranker queueing because GPU capacity is not elastic on a five-minute horizon. Systems that survive are designed to shed — precomputed ranking for the head, a BM25-only degrade path, and a load-shedding order decided in advance rather than at 00:04 on sale night.
The catalogue is the ceiling
Search, recommendations, forecasting and pricing all read the same product record. If the same shoe is listed fourteen ways across sellers, with size in three formats and colour in free text, no amount of model quality recovers it. Measuring attribute coverage and accuracy before any ranking work is not a delay — it is the only way to know whether a relevance problem is a relevance problem.
The search path has a hard latency budget
A shopper-facing search is a few hundred milliseconds end to end, and that budget is already spent on the storefront, the API gateway and the facet counts. Every component has to justify its slice: an embedding call at query time is a network hop you own, and a cross-encoder over two hundred candidates will not fit. The architecture question is which model can run inside the budget, not which model scores best offline.
Inventory truth is distributed and stale
Available-to-promise spans a distribution centre, stores, dark stores and in-flight transfers, each with its own update cadence and its own lie. Oversell is a refund and a complaint; over-reservation is margin left on the shelf. Reservation semantics, a reconciliation loop and an explicit staleness budget per channel are load-bearing design work, because a forecast built on a fictional on-hand number is a fiction with a confidence interval.
Margin decides what inference is allowed
In a category running single-digit contribution margin, a large model on every search query is not a feature, it is a tax. The usual resolution is tiering: distil the head of the query distribution into a small model, cache semantically, batch what can be offline, and reserve the expensive path for the tail and for high-intent moments. Cost per thousand queries belongs on the same dashboard as relevance, or it will only be discovered in the monthly invoice.
Every one of these is discoverable in week one. Most programmes discover them in month three, after the architecture is already load-bearing.
Outcomes
What we design toward.
Targets and typical ranges for this class of system. Each one carries its basis, because a number without one is decoration.
- Attribute coverage reachable on a messy marketplace catalogue
- 60–85%Attribute coverage reachable on a messy marketplace catalogueTypical range for vision-plus-text extraction with a confidence gate and a human review queue. The ceiling is set by supplier image and spec-sheet quality, not by the model — we baseline it on a sample of your own catalogue before committing to a number.
- p99 design target for the full hybrid search path
- < 250 msp99 design target for the full hybrid search pathDesign target we architect and load-test against — retrieval, fusion and rerank inside the budget, with a lexical-only degrade path when the reranker is saturated. Not a measured result on your infrastructure until we have measured it.
- Peak-to-baseline traffic multiple the platform is sized for
- 10–40×Peak-to-baseline traffic multiple the platform is sized forRange consistent with publicly reported festive-sale and Black Friday peaks for large marketplaces. We size against your own worst hour from the last three years, plus headroom, rather than against this figure.
- Inference cost reduction available on the head of the query distribution
- 5–20×Inference cost reduction available on the head of the query distributionTypical range when a distilled small model plus a semantic cache absorbs repeated queries. Measured per system; the tail still routes to the large model, so blended savings are lower than the headline.
These are design targets and published benchmark ranges, not delivered client results. We do not publish results attributed to clients.
Solutions
What we bring to this sector.
The practices that do most of the work here. Each one is the same discipline applied against the constraints above.
Blueprints
Reference architectures behind this work.
How the systems are actually assembled, with the engineering decisions and their costs stated. Reference designs, not case studies.
Grounded Answer Engine
A question-answering system over a company’s own documents that cites every claim to a source span, respects the permissions of the person asking, and refuses when the evidence is not there.
Read the blueprintReal-Time Decision Platform
A platform that turns database changes into fresh features within seconds, serves a model decision inside a live request, and writes an immutable record of every decision — the inputs, the model version and the policy that applied — so it can be explained, replayed and rolled back months later.
Read the blueprintQuestions
The awkward questions.
The ones a technology committee actually asks, answered the way we would answer them in the room.
Next step
Bring us the constraint you keep working around.
A 45-minute working session. We'll tell you what we'd build, what we'd not build, and roughly what it costs. No deck.