Kenny Ponders.

KENNY WU · AI ENGINEER · SHENZHEN

From AI prototypes
to reliable delivery.

I’m Kenny, an AI systems and application engineer.
I build with language, retrieval, and multimodal models, turning raw model capability into dependable workflows.

01 / SELECTED DELIVERY

Professional work · Roles & outcomes

· Technical lead

AI comics generation

Generation is only step one.
Automated evaluation and self-healing make it work.

Orchestrated end-to-end pipelines across script analysis, asset extraction, and multi-shot storyboarding, while designing automated evaluation and acceptance gates.

See responsibilities and outcomes ↗
From script to evaluated storyboardsWorkflow diagram
  1. 01 / UNDERSTANDUnderstand scriptsIdentify characters and plot
  2. 02 / ORGANIZEExtract assetsOrganize characters and scenes
  3. 03 / GENERATEGenerate shotsStructured shot decomposition
Independent quality checks→Budgeted retry / auto-fix→Acceptance

Runtime safeguards: circuit breaking · concurrency management · automated self-healing

Issue rate · Fixed evaluation set

≈18%→9.53%

Action overload & shot-scale issues · ≈47% relative reduction

One real-world run

20 episodes

Completed end-to-end with automated error recovery

2024 / RETRIEVAL & ANSWERING

Enterprise document Q&A

Paired hierarchical document chunking with hybrid multi-path retrieval and dynamic thresholds to deliver accurate, hallucination-resistant enterprise Q&A.

See this work ↗
Document structureMulti-path retrievalAnswer evaluation
≈90%

Internal tests
Recall and Q&A accuracy

02 / ENGINEERING JUDGMENT

How I move the work forward

Questions I return to ↗
01

Define "good" before building.

Turn subjective quality into quantifiable metrics. Use a frozen evaluation set to verify whether every change actually helps.

02

Design with failure as a given.

LLM calls inevitably hiccup. Build checkpoint recovery and retry limits so isolated failures never break the entire pipeline.

03

Focus on the full pipeline, not just single generations.

Real-world delivery is won in preprocessing, state handling, concurrency, and graceful fallbacks around the model.

03 / OPEN-SOURCE WORK

Turning ideas into everyday, crafted tools.

All projects ↗

04 / FIELD NOTES

Field notes

All notes ↗
How does Perplexity design agent skills?↗Agent context engineering: lessons from Manus↗Blind spots in agent search systems↗

LET’S WORK TOGETHER

Planning or scaling an AI project?

Whether you're exploring architecture, solving production delivery, or tuning stability, feel free to reach out.

jdlow@live.cn ↗