← CC-Analysis

Pipeline Enhancement Ideas

Force-multipliers and gap-fillers for the TroveSnap pipeline — beyond what's already planned. Each idea is tagged for how it fits the candidate-first, deterministic-first, cost-controlled architecture. Brainstorm, 2026-06-17.

1 Capture 2 Review 3 Intelligence 4 Orchestration 5 Outbound 6 Trust

Ideas are arranged along the existing pipeline: external source → candidate → review → canonical → publish → demand → disposition.

Feasibility: deterministic ships keyless / free needs model AI; some free-capable (local) token-metered premium AI, paid path needs Auth / connectors gated on later phases ★ top pick

If we pick three to pursue first

Highest leverage against the core pain, the candidate layer, and the competitive wedge.

★ Voice cataloging

Hands-free intake during the walkthrough — the single biggest attack on the 14–19 min/lot pain.

★ Cross-run dedupe

Makes the whole candidate layer trustworthy at scale; links reappearing items instead of cloning.

★ PII / privacy scrubber

Turns a competitor's CCPA weakness into TroveSnap's headline trust feature.

Channel-export adapters are the highest value but gate on Auth + connectors — a phase-2 bet.

1 Capture — front of the pipeline

Get items in faster and richer, before the form ever appears.

★ top pick

Voice cataloging on the walkthrough

Photographer speaks "oak dresser, 1960s, ask $120"; transcription pre-fills candidate fields paired to the photo. Hands-free.

needs modellocal Whisper = free

Video-sweep → frame sampling

Walk a room on video; sample frames, detect distinct objects, emit candidates. One pass catalogs a whole room.

needs model

Single-photo decomposition + lot bundling

A shelf photo becomes N draft items via segmentation; the inverse recommends grouping low-value items into one sellable lot.

needs model

Provenance capture lane

Photograph receipts, COAs, appraisal letters → attached as evidence to the item, boosting trust/value and feeding the appraiser.

deterministicOCR

2 Candidate / review — where quality is won

Reduce review noise and speed throughput without ever bypassing the promotion contract.

★ top pick

Cross-run + cross-sale dedupe

Perceptual hash + embeddings detect the same item reappearing (relisted, or duplicated across connectors) and link instead of cloning.

needs modelcheap / local

Confidence-tiered promotion policy

Seller rule: high-confidence, low-value, clearly-identified items auto-promote; uncertain or high-value always hold. A policy over promoteInventoryCandidate(), not a bypass — auditable.

deterministic

Active-learning review

The queue learns each seller's repeated edits (always recategorizes "misc") and pre-fills/reorders accordingly. Enrichment that compounds.

needs model

3 Intelligence / enrichment

Make guidance trustworthy and self-improving.

Comparable-evidence cards

Pricing guidance shows the actual sold comps behind the range with recency/condition adjustments, so sellers trust it.

surface freedeep comps metered

Outcome feedback loop

Reconciled sold prices calibrate future guidance per category/region. Guidance → sale → reconciliation → better guidance — the learning flywheel.

deterministic

Authenticity / red-flag detector

Flags likely reproductions in high-value categories to trigger closer appraisal. Risk reduction for client estates.

needs model

4 Orchestration — the layer agencies pay for

Turn the Kanban into an operations tool.

Pipeline throughput + SLA dashboard

Time-in-stage, bottleneck detection ("40 items stuck in Needs Appraisal 3 days"), per-cataloger throughput.

deterministic

Assignable tasks + roles

Assign review / appraisal / reshoot work to teammates (device tokens already exist). Multi-user cataloging.

deterministicneeds Auth

Visible provenance timeline

Surface each item's full lineage — photo → run → connector → edits → promotion — already tracked via promoted_record_id. Trust + compliance artifact.

deterministic

5 Outbound — "catalog once, list everywhere" made real

Project the canonical record outward as drafts — never auto-posted.

Auction-pipeline outbound

Per-target export profiles for the feeds estate sellers use — LiveAuctioneers, Proxibid, Invaluable, Auction Flex/HiBid, Wavebid — each with its own template + lot-numbered image bundle, plus sale-results pull-back. See the integration spec.

file export earlierAPI needs Auth

Channel-shaped export adapters

Per-channel resale listing payloads (eBay item specifics, FB fields, Shopify) from the canonical record — drafts, candidate-style. Makes the reseller promise shippable.

deterministicneeds connectors

Physical bridge

Print price tags / QR item labels (QR → buyer item page + door code) and a printable catalog from the same records. Connects the sale floor to the pipeline.

deterministic

6 Trust — the sharpest competitive wedge

Make safety a headline feature where incumbents are exposed.

★ top pick

PII / privacy scrubber at ingestion

Auto-detect faces, addressed mail, documents with SSNs, people in reflections → flag/blur before anything publishes. Directly targets the EstateSales.net interior-photo / CCPA complaint we found, and reinforces "nothing publishes without approval."

needs model
How to read the tags. "Deterministic" ideas can ship keyless and stay green in CI (the project's baseline rule). "Needs model" ideas have free-capable local paths (e.g. Whisper, perceptual hash) or escalate to token-metered premium. "Needs Auth / connectors" ideas are gated on later roadmap phases — sequence them after the free-capable wins. None of these change the ingestion contract: everything still flows source → candidate → review → promoteInventoryCandidate().