Data plane / Training studio
BM
Data plane · active learning (B6)

Training studio

Claude labels → ML.NET serves at scale → low-confidence cases feed the loop
M1 · Condition classifierserving
new · like-new · good · fair · poorv1.7
Accuracy94.2%
Labels18.4k
Retrain 24h+412 this week
M2 · Feature extractorserving
sealed · gatefold · 180g · OEM · with_casev2.1
Accuracy91.5%
Labels22.9k
Retrain 24h+536 this week
M3 · Product matcherserving
cross-venue identity · gray-zone pairsv1.3
Accuracy88.9%
Labels9.1k
Retrain 24h+287 this week
M4 · Brand / Categoryserving
brand + category from titlev3.0
Accuracy96.1%
Labels31.6k
Retrain 24h+624 this week

Active-learning loop

looping · hot-swap
Sample low-confidence Claude labels Grow seed set Retrain Serve
💡
The ML.NET classifiers serve at scale; only cases below the confidence threshold escalate to Claude for labels (ConditionLabel / FeatureSynonym / ProductMatchLabel). New labels grow the seed set, the model retrains every 24h and hot-swaps in.
TypeExample titleLabelDecided byWhen
ConditionLabelSony WH-1000XM4 — open box, testedlike-new🤖 Claude2m ago
FeatureSynonymMiles Davis Kind of Blue 180 gram audiophile180g🤖 Claude14m ago
ProductMatchLabel"WH1000XM4" ↔ "WH-1000 XM4 headphones"match🤖 Claude31m ago
ConditionLabelCanon G7X II - some scratches, works greatgood🤖 Claude48m ago
FeatureSynonymRolex Sub w/ box and papers full setwith_case👤 Operator1h ago
ProductMatchLabel"Prusa i3 MK3S" ↔ "Original Prusa MK3S+"no-match👤 Operator2h ago

Confirmed field mappings per source

Training studio · field mappings per source
SourceFieldMapping / selectorConfidenceStatus
eBaytitle$.Item.Title → normalizedTitle0.98confirmed
eBaycondition$.Item.ConditionID → condition enum0.95confirmed
Mercariprice.mer-price__value → priceCents0.91confirmed
Mercaribrand.item-brand a → brandRaw0.79review
B-Stockmanifesttable.manifest tr → lotItem[]0.84confirmed
B-Stockquantity.lot-qty → unitCount0.67draft

Training studio · Phase 3 — self-serve site config

Planned

A non-engineer adds a new site config without code: LLM-assisted JSON mapping plus point-and-click DOM annotation, with a Tier-2 CSS / XPath fallback when structured feeds aren't available. Extends the field-mapping table above to net-new sources.