Technical Overview
rScan is a production multimodal system, not a research demonstrator. This page covers where inference runs, what each stage produces, and the engineering posture that keeps a learned component from ever putting a live listing into a wrong state. It is the overview we share with technical reviewers; deeper implementation detail is available under NDA.
System Overview
Each stage hands a typed record to the next. A learned stage that fails or returns low-confidence output degrades to the deterministic path beneath it — the unit still catalogs, with less enrichment. No inference failure can block receiving.
Stage Detail
Decode runs in the capture client, directly against the device camera, and resolves UPC, EAN, GTIN and Amazon FNSKU identifiers locally. It is the one stage that sits between an operator and their next physical action, so it never waits on the network.
Every image entering the platform is deduplicated, checked for authenticity — genuine unit photography versus manufacturer stock renders — and bound to the correct product family with a provenance record. Attributes an identifier cannot supply, such as condition and configuration, are resolved by multimodal inference over imagery and recovered text.
A provider-portable inference layer turns unstructured inputs into a schema-validated product record, classifies it against the Google Product Taxonomy, canonicalizes the brand, and applies each marketplace's policy rules before publication.
Price selection is a learned policy trained on rScan's proprietary corpus of more than 1.78 million secondary-market observations. Comparable quality is enforced ahead of the model, and every price is bounded by a computed cost floor.
Publication targets five first-party marketplace integrations, with quantity reconciled continuously so one unit is never sold twice. Concurrent changes are serialized, and every create, revise and end is written to an auditable ledger that outlives the process that made it.
AI Capability Map
Model selection is per-task and provider-portable. Decisions with a hard correctness requirement use validated structured output; advisory decisions run under a latency budget and fail open.
| Capability | Input | Output | Approach |
|---|---|---|---|
| Identifier recovery | Live camera frame | UPC / EAN / GTIN / FNSKU | On-device decode |
| Image integrity | Product imagery | Dedup, authenticity, provenance | Computer vision |
| Attribute extraction | Imagery + unstructured text | Schema-validated record | Multimodal LLM |
| Taxonomy & brand | Catalog record | Category node, canonical brand | LLM classification |
| Comparable matching | Market candidates | Verified comparable set | Scoring + model review |
| Valuation | Market observations | Price action | Reinforcement learning |
Proprietary Data Assets
These datasets are generated as a byproduct of operating real liquidation inventory at volume. They are not licensed, resold, or synthetic.
Price observations
1.78M
Secondary-market comparables with capture time, matched product and market context. Direct input to the valuation policy.
Operator event stream
1.16M
Structured physical-workflow events — scanning, stocking, picking, audit, returns — the labeled record of how a unit actually moves.
Catalog templates
77,116
Normalized product records with attributes, taxonomy and vetted imagery, reused so repeat units resolve instantly.
Canonical brands
3,862
Manufacturer brands normalized from free-text source variants across the catalog.
Compute Profile & GPU Roadmap
What runs where today, and the three workloads where dedicated GPU capacity moves rScan from orchestrating hosted inference to owning it.
Identifier decode runs on the operator's device. Image processing, catalog synthesis and taxonomy classification run server-side, with the language and vision stages served by hosted frontier models under per-call latency and quota budgets. The valuation model trains on GPU-accelerated TensorFlow and serves on CPU, which is adequate at the current decision rate.
Engineering Posture
Every learned component sits on a deterministic foundation. An inference failure degrades a record; it never blocks receiving or produces a wrong listing unnoticed.
Every marketplace mutation is written to a durable event ledger that can reconstruct what was attempted and what the provider answered, long after the process has exited.
Revenue-bearing operations always take priority over background observation. Learning never slows a sale.
Model providers sit behind one interface, so a model upgrade is a configuration change and no single vendor is a dependency.
We are glad to walk an engineering audience through the platform in depth, including the parts that are deliberately still deterministic and why.
