Rina
Agustina

A seller lists a generic phone case but embeds brand names like "Samsung," "Apple," and "Xiaomi" in the product title and description — despite the item not being compatible with or affiliated with those brands. The tactic inflates search visibility and misleads buyers.

1. Cross-referenced product title/description against brand taxonomy rules.
2. Checked if claimed brand relationship was verified in the seller's brand authorization records.
3. Reviewed listing images and specs for product-brand alignment.
4. Flagged and escalated with evidence tagging: "Unauthorized brand mention," "Misleading title."
5. Applied category-appropriate action per policy matrix (listing removal / warning).

Listing removed within SLA. Seller received first policy strike. Case documented for potential pattern-based review in future audits. Zero false positive rate in sampled calibrations for this abuse type.

Draw precise bounding boxes around objects in product images for a computer vision model training dataset. Each image required identifying multiple object classes (product, packaging, label, background clutter) and labeling them according to a taxonomy with 20+ categories.

1. Read and internalized detailed annotation guidelines before starting each task batch.
2. Applied consistent bounding box tightness (no padding rule, edge-touching preferred).
3. Used zoom-in verification for ambiguous or overlapping objects.
4. Self-QA'd 100% of outputs before submission.
5. Flagged edge cases with notes for reviewer adjudication.

High sustained attention to detail, fast guideline adoption, self-correction without reviewer dependency, high throughput with quality maintained.

A seller originally lists a genuine branded item at a competitive price, gains reviews and ranking, then silently updates the listing to ship a counterfeit or significantly inferior product — a "bait-and-switch" tactic. The original listing details are retained but fulfilled differently.

1. Pulled switched listing reports from internal moderation platform.
2. Compared original listing snapshot vs. current listing across: title, description, images, price, specs, and seller SKU.
3. Scored change severity on a 3-tier matrix (Minor Edit / Significant Change / Policy Violation).
4. For violations: documented evidence fields, referenced the relevant policy clause, and escalated for actioning.

Identified 15–30 switched listing violations per day on average. Escalated cases contributed to seller warning, suspension, or listing removal pipelines, protecting buyer trust and platform integrity.

Brand verification cases arrive from two sources: brand owner complaints and proactive sweep queues. Each case requires reviewing whether a seller's listing legitimately represents a brand or constitutes IP misuse — including counterfeiting, unauthorized use, and misleading brand association.

1. Check seller's brand authorization status in verification registry.
2. Audit listing for: logo misuse, trademark in title, misleading "official" claims, fake certificate images.
3. Assess product imagery for authenticity signals (genuine vs replica packaging).
4. Classify under IP violation type: Counterfeit / Unauthorized Use / Misleading Association.
5. Apply action per brand owner's enforcement preference (remove / delist / warn).

Trademark recognition, policy interpretation, high-volume case management, fast decision-making with low error rate, cross-referencing multiple data sources simultaneously.

Label text samples (customer reviews, chat messages, social posts) across multiple dimensions for NLP model training: sentiment (positive/neutral/negative), intent (query/complaint/compliment/spam), toxicity level (0–3 scale), language (Bahasa/English/mixed), and topic category.

1. Read full text before labeling — avoided first-impression bias.
2. Applied multi-label logic for ambiguous samples (e.g., sarcastic positive = neutral sentiment + complaint intent).
3. Maintained a personal edge-case log for consistency across sessions.
4. Leveraged native Bahasa Indonesia fluency for mixed-language content.
5. Achieved 97%+ inter-annotator agreement score in calibration rounds.

Native Bahasa Indonesia fluency was a significant quality advantage for Indonesian-language datasets, reducing ambiguity flags and improving label precision for local-language nuances.

Marketplace catalog integrity depends on products residing in correct category nodes. Miscategorized products reduce search relevance, hurt buyer experience, and expose sellers to policy violations. This task involved validating and correcting category assignments at scale.

1. Reviewed product title, images, description, and attributes.
2. Matched product characteristics against the platform's 3-level category taxonomy (L1 > L2 > L3).
3. Identified miscategorizations (e.g., clothing listed under Beauty, electronics in Sports).
4. Applied correct L3 category assignment with justification note.
5. Flagged borderline cases for policy team adjudication.

Maintained 99% category assignment accuracy across 800+ daily reviews. Contributed to improved search relevance scores and reduced buyer mismatch complaints in the moderated category segments.