Chen Canxu

Participated in data construction and AI training data projects for electronic accessories sales and after-sales scenarios, covering customer-service QA and RAG evaluation, e-commerce after-sales image recognition and issue-attribution evaluation, multimodal material screening, image/video structured annotation, and image-understanding model response review.

• Organized customer-service QA, historical conversations, and knowledge-base materials into structured RAG evaluation samples, aligning user questions, retrieved materials, and model answers.
• Evaluated whether answers matched user intent and business rules, identifying missing key information, insufficient evidence, irrelevant answers, unsupported expansion, and inconsistent customer-service wording.
• Annotated abnormal samples by root cause, rewrote selected reference answers, and summarized high-frequency issues and missing knowledge points to support FAQ and knowledge-base optimization.
• Organized and evaluated about 5,000 e-commerce customer-service QA and RAG samples across product consultation, specification matching, after-sales policy, logistics exceptions, refund rules, and after-sales handling.
• Evaluated about 7,000 after-sales image recognition and issue-attribution samples, covering damaged goods, packaging issues, wrong or missing items, accessory omissions, specification or color mismatch, and insufficient evidence.
• Annotated and reviewed about 1,500 multimodal samples, including image/video quality screening, visual attributes, and image-understanding model responses.

Worked on Cantonese ASR transcription normalization, converting Cantonese spoken expressions into natural and fluent Mandarin text while following project annotation standards.

• Normalized Cantonese ASR initial transcripts into standard Mandarin, correcting dialect words, colloquial expressions, word-order differences, incomplete semantics, and unnatural phrasing based on context.
• Completed text proofreading and consistency checks according to annotation guidelines, and fed back common conversion issues to improve data usability.
• Completed about 9,000 Cantonese ASR normalization and proofreading records, helping improve the accuracy, consistency, and training usability of dialect speech transcription data.

Structured user questions, retrieved materials, and model answers into evaluation samples, then assessed intent matching, evidence sufficiency, answer completeness, unsupported expansion, and customer-service policy consistency.

• Covered product functions, specifications, logistics exceptions, shipping timeliness, return/exchange rules, refund rules, and refund timeliness.
• Summarized reusable issue categories such as acceptable answers, answers needing rewrite, missing materials, inaccurate retrieval, unclear intent, and inconsistent wording.

Built a Dify Workflow-based evaluation assistant with separate evaluation-rule and after-sales knowledge bases, standardizing review of user questions, retrieved materials, and model answers.

• Output four pre-review suggestions: acceptable, needs rewrite, unacceptable, and manual review, while generating error attribution, judgment basis, rewriting suggestions, and QA notes.
• Built a local HTML console for single-case evaluation, CSV batch processing, golden-set validation, issue statistics, and result export, with desensitization, version checks, and human confirmation.
• Applied to real RAG evaluation work, improving initial review and retrospective organization efficiency by more than 50% while keeping high-risk samples under manual review.

Evaluated model recognition accuracy and issue attribution quality for electronic accessories after-sales scenarios using user-uploaded images, issue descriptions, order information, and model outputs.

• Reviewed judgments for product damage, packaging damage, missing accessories, wrong or missing shipments, color/model mismatch, obvious usage traces, and insufficient image evidence.
• Recorded model output errors and boundary cases, including recognition errors, omissions, false positives, unreasonable attribution, insufficient evidence, blurry images, occluded angles, and mismatches between text and image.
• Helped consolidate review rules for image recognition errors, attribution bias, evidence sufficiency, and QA judgment.

Screened image and video materials, annotated visual attributes, and evaluated image-understanding model responses for factual accuracy, completeness, format compliance, and instruction following.

• Screened materials by clarity, exposure, watermark, subject completeness, frame stability, and content usability.
• Classified images into photography and illustration categories, classified videos into live-action, animation, and mixed video, and recorded reasons for rejection or uncertainty.
• Annotated tone, saturation, camera angle, composition, lighting type, shot scale, and camera movement, while supplementing or rewriting human reference answers when needed.