ABOUT / KENNY WU

Between research and practice,
accountable for the outcome.

I’m Kenny Wu, an AI engineer based in Shenzhen. My work has grown from natural language processing and information retrieval to multimodal understanding and content generation, with a focus on turning model capabilities into useful products.

Contact me ↗Google Scholar ↗

≈18% → 9.53%

Comics · Issue rate on a fixed evaluation set
Action overload and shot-scale issues

1285s → 24s

Short drama · One 10-episode sample
Identity disambiguation time

≈90%

Enterprise Q&A · Internal tests
Recall and Q&A accuracy

Experience

2026

AI comics & storyboard generation

Technical lead

Helped shape the system architecture and core workflow, connecting script understanding, asset extraction, and storyboard generation. Led quality evaluation and acceptance criteria, with recovery mechanisms for long-running generation tasks.

Fixed evaluation set: action-overload and shot-scale issues fell from about 18% to 9.53%. One 20-episode run completed despite intermediate failures.

Work & outcomes

Built structured storyboards and shot decomposition, with independent quality checks and bounded correction.
On a fixed evaluation set, action-overload and shot-scale issues fell from about 18% to 9.53%—a relative reduction of about 47%.
Added circuit breaking, concurrency control, and bounded recovery; a 20-episode run completed despite intermediate failures.

2025–2026

Short-drama editing & character identity

Workstream lead

Designed cross-episode identity, entity alignment, and highlight retrieval for short-drama editing. Connected character understanding to the editing workflow so footage could follow consistent character threads.

One 10-episode sample: identity disambiguation fell from 1,285s to 24s; understanding orchestration fell from 23 minutes to about 5–6 minutes.

Work & outcomes

Improved data services, concurrent scheduling, and cache reuse to support the editing pipeline.
On one 10-episode sample, identity disambiguation fell from 1,285 to 24 seconds; understanding orchestration fell from 23 minutes to about 5–6 minutes.

2025

E-commerce copy style adaptation

Core algorithm engineering

Built a style-adaptation workflow for product narration, spanning dataset creation, model training, and evaluation for digital-human content production.

Fine-tuned Qwen2.5-7B for five copy styles.

Work & outcomes

Designed a source-text → style-profile → rewrite dataset workflow and multidimensional automatic evaluation.
Fine-tuned Qwen2.5-7B for five copy styles, using structured style information to stabilize training.

2024

Enterprise RAG, safety & intent recognition

Core algorithm engineering

Worked on enterprise document Q&A, internal LLM safety review, and voice-bot intent recognition, covering retrieval, generation, and interaction quality.

Internal tests reported about 90% recall and Q&A accuracy.

Work & outcomes

Combined paragraph hierarchy, multi-path retrieval, and dynamic thresholds; internal tests reported about 90% recall and Q&A accuracy.
Combined semantic matching and rules for content review, supporting the internal model registration process.
Used Aho–Corasick matching and inverted indexes for intent recognition across scenarios.

2022–2024

Intelligent search & conversation analysis

NLP algorithm trainee

Worked on intelligent search, Text2SQL, and social-media support conversations, covering query understanding, similar-question grouping, and discovery of unanswered questions.

Work & outcomes

Generated SQL templates from database schemas and plain-language explanations for business review.
Used contrastive sentence embeddings and hierarchical density clustering to expand question libraries and support conversation reporting.