冯靖婉 AI 落地工程师 / Forward Deployed Engineer Coco Feng AI Solutions Engineer · Forward Deployed
交互设计出身，游戏制作 × 产品设计 × AI 应用的跨领域背景。
自驱探索 AI 工程，已产出多项团队实际使用的提效工具。
8 年加拿大留学与工作经历，英语流利，适应跨文化协作。
务实导向，理解制作管线痛点，关注技术与工具协同后的场景落地。
Interaction design background — game production × product design × AI applications.
Self-directed in AI engineering, with multiple productivity tools already in use by the team.
8 years of study and work in Canada — fluent English, comfortable in cross-cultural collaboration.
Pragmatic, grounded in production-pipeline pain points; focused on where tech and tools actually land in real scenarios.
SFU 互动艺术与技术
SFU Interactive Arts
外包供应商管理
Project coord · Vendors
本地模型 · 开源选型
Local models · OSS
英语流利
Fluent English
交付导向。从需求洞察、工具选型、Workflow 设计到交付验证独立走完——评判一件事是否做完，标准是「能用、可复用」。同时创立觉然，做 AI 方向的多元探索。
Delivery-first. Owns the loop end-to-end — discovery, tool selection, workflow design, delivery validation — with "actually usable, reusable" as the bar. Also founded JORAN as an ongoing space for exploration across AI directions.
- 自驱 AI 工程实践：在职责外自学 AI Workflow，产出多项面向制作环节的提效工具
- 剧本审核 Workflow：Google AI Studio 搭基础流水线，叠加 Claude / Antigravity 做多步骤批处理，对百万字游戏剧本完成批量校对与风险核查；交付内容经人工验证 100% 通过
- CG 参考图 Workflow：替代画师传统的「找参考 → 画线稿」流程，AI 批量生成人体结构、交互动作等方向的参考图，画师只需在产出中筛选可用图片
- 配乐参考 Workflow：调研对比 ACE-Step 1.5 等开源音乐生成模型后选 MIDI 程序化生成路线；设计 Leitmotif 主导动机系统，用 Python midiutil 批量生成 44 首 MIDI，经 FluidSynth + GeneralUser GS SF2 渲染为 MP3/WAV，作为项目配乐参考
- 本地模型基建：Hugging Face 选型 + Ollama / LM Studio 本地部署 + Cloudflare 反向代理保障开发访问，搭出可独立运行的推理环境
- 项目管理：负责排期与资源分配，保障任务按时交付；协助管理外包供应商流程
- AI 技术架构顾问（2026.03 起）：转为顾问角色，为公司 AI 技术选型与 Workflow 集成方案提供指导；同期创立觉然 JORAN，探索 AI 教育方向
- Self-directed AI engineering: Picked up AI workflow engineering outside production duties; shipped multiple productivity tools for production stages
- Script Review Workflow: Three-stage pipeline (Google AI Studio + Claude + Antigravity) handling million-character batch proofreading and risk checks on game scripts; 100% pass rate on manual verification
- CG Reference Workflow: Replaces the artist's traditional "search references → draft line art" loop — AI batch-generates references for anatomy and interaction angles, artists pick what they need from the output
- Music Reference Workflow: After evaluating ACE-Step 1.5 and other OSS music generators, chose programmatic MIDI; designed a Leitmotif system, batch-generated 44 MIDI tracks via Python midiutil, rendered through FluidSynth + GeneralUser GS SF2 to MP3/WAV as project music references
- Local Model Stack: Hugging Face evaluation + Ollama / LM Studio local deployment + Cloudflare reverse proxy for dev access; built a self-sufficient inference environment
- Project management: Owned scheduling and resource allocation, on-time delivery; supported vendor management
- AI Architecture Consultant (from Mar 2026): Transitioned to advisory role on AI tech selection and workflow integration; concurrently founded JORAN to explore AI education
- 参与《侏罗纪世界：进化3》（已发布）等国际游戏项目 VO/ADR 录音工作
- 严格执行数字资产管理规范，在国际化制作流程标准下保障交付质量
- VO/ADR recording for Jurassic World Evolution 3 (shipped) and other international game titles
- Strict digital asset management within an international game production pipeline
- 端到端设计执行：需求分析 → Figma 高保真 UI → 设计系统搭建
- 与产品经理和工程师协作，输出设计规范文档和部分 PRD
- End-to-end design ownership from requirements through high-fidelity Figma UI delivery; built and maintained design system
- Collaborated with PMs and engineers; produced design specs and contributed to PRD documentation
- 参与品牌焕新与官网重建，协助从策略到视觉执行的落地
- 统筹 Art Vancouver 2022 策展协助与现场执行，协调多方资源
- Contributed to brand refresh and website rebuild — supporting strategy through visual execution
- Coordinated Art Vancouver 2022: curation support, on-site execution, artist and logistics management
这些项目按时间排，跨度差不多半年。AI 工具变化太快——下面每个案例里的组合，都是当月我跑通、用熟、能交付的选型。换个时间点再做，我会换一套，这正是这份工作有意思的地方。
These projects span about half a year, in chronological order. AI tools move fast — the stack you see in each case is what I evaluated, picked, and shipped with that month. If I were doing it today I'd reach for something different — that's part of what makes this work fun.
背景 · Rejet 上海工作室游戏剧本规模达百万字级别，原本依赖人工逐章校对与风险词审核，效率瓶颈明显，且合规敏感度高。
挑战 · 体量大；风险词识别需结合上下文；角色一致性贯穿全文；输出格式必须能直接回传产线。
方案 · 三段流水线 Workflow：① Google AI Studio 搭基础校对节点；② Claude / Antigravity 多步骤批处理读取整目录文件 1–N，做内容核查 + 风险标注；③ 结果回写为同格式文档。
价值 · 交付内容经人工逐项验证 100% 通过；将"逐章人工审"压缩为"批量审 + 人工抽样"。
Background · Game scripts at Rejet's Shanghai studio reach the million-character scale. Manual chapter-by-chapter review was the bottleneck, with high compliance sensitivity.
Challenges · Sheer volume; risk-word detection needs context; character consistency must hold across the full text; output must drop straight back into the pipeline.
Solution · A three-stage workflow: ① Google AI Studio for the base review node; ② Claude / Antigravity for multi-step batch processing across files 1–N (content check + risk flagging); ③ Results written back into the original document format.
Impact · 100% pass rate on manual verification; review shifted from "every chapter by hand" to "batch review + sampling."
背景 · 游戏 CG 绘图前期，画师传统流程是先手动搜寻人体结构、人物互动姿态等参考资料，再画线稿展开后续绘制。该环节耗时大、参考质量参差。
挑战 · 把"人体结构 / 交互动作"等抽象需求拆解为可批量生成的视觉关键词；产出多维度参考图，质量稳定到画师能直接用。
方案 · 针对画师手动搜寻人体结构、交互动作的环节做工具化优化，用 AI 批量生成对应方向的参考图，画师只需在产出中筛选可用图片即可。
价值 · 把"画线稿前先搜参考"的传统流程替换为"批量出图 + 筛选"，画师在 CG 绘图前期的参考准备时间显著缩短。
Background · In CG pre-production, the traditional artist workflow is: manually search references for anatomy, pose, character interaction → draft line art → continue. Reference search alone consumes hours and quality varies.
Challenges · Translate abstract needs ("anatomy / interaction") into batch-generatable visual prompts; ensure the output is consistent enough for direct artist use.
Solution · Tool-driven optimisation of the artist's reference-hunting step: AI batch-generates reference images for anatomy and interaction angles; artists pick what they need from the output.
Impact · Replaces "search references → draft" with "batch-generate → filter," cutting the reference-prep time at the front of CG production.
背景 · 游戏配乐立项阶段需要大量风格化参考音轨，外部采买成本高、内部沟通"想要什么风格"的成本也高。
挑战 · 开源音乐生成模型（如 ACE-Step 1.5）质量不稳定且不可控；需要一套可批量、能调风格、能体现剧情主题的统一音乐语言。
方案 · 调研对比后选 MIDI 程序化生成路线 → 设计 Leitmotif 主导动机系统统一音乐语言 → Python midiutil 批量生成 → FluidSynth + GeneralUser GS SF2 渲染 MP3/WAV。
价值 · 一次交付 44 首 MIDI 参考音轨，作为项目制作参考被采纳；同事和团队在选向、定调阶段直接使用。
Background · Music pre-production needs lots of style references — buying externally is expensive, and aligning internally on "what style do we want" is just as costly.
Challenges · OSS music gen (e.g. ACE-Step 1.5) was unstable and not controllable; we needed a batchable, style-tunable, theme-coherent musical vocabulary.
Solution · After evaluation, chose programmatic MIDI → designed a Leitmotif system for thematic coherence → Python midiutil for batch generation → FluidSynth + GeneralUser GS SF2 for MP3/WAV rendering.
Impact · One-shot delivery of 44 MIDI reference tracks, used as the project's music reference for direction-setting.
背景 · 个人聊天记录是日常最自然的语料，但默认加密、格式封闭，AI Agent 无法直接读取与处理。
挑战 · 需要解密本地数据库、处理多端结构差异；同时把"读取 / 检索"以 MCP 标准协议暴露出去，让任意支持 MCP 的 Agent 都能调用。
方案 · 用 FastMCP 实现一个微信数据解密 MCP Server，把"读取聊天 / 检索关键词 / 提取片段"封装为标准 MCP 工具，Agent 通过自然语言即可调用。
价值 · 一套可复用的本地数据 MCP 工具链；让 AI Agent 在聊天记录上做总结、检索、分析等任务，不再依赖手动复制粘贴。
Background · Personal chat history is the most natural daily corpus, but it's encrypted and locked in proprietary formats — AI agents can't read it directly.
Challenges · Decrypt local databases, handle structure differences across clients, and expose "read / search" as a standard MCP protocol any agent can call.
Solution · Built a WeChat decryption MCP server with FastMCP — wraps "read chats / keyword search / extract snippets" as standard MCP tools that agents invoke through natural language.
Impact · A reusable local-data MCP toolchain; agents can summarise / search / analyse chat history without manual copy-paste.
背景 · 个人项目，加密货币永续合约的自动交易系统。
挑战 · 加密货币市场噪音大、单信号容易过拟合；需要一套可观测、可灵活调参、能稳定执行的整体系统。
方案 · 多信号评分 Workflow（EMA / RSI / 布林带 / 波动率 / 情绪）驱动开仓决策；自适应止盈止损 + 追踪止盈；实时仪表盘监控。
价值 · 跑通从"信号输入 → 决策 → 执行 → 风控"的完整 Workflow，所有节点可独立观测与调参。
Background · Personal project: an automated trading system for crypto perpetual futures.
Challenges · Crypto markets are noisy; single signals overfit fast; the whole system has to be observable, retunable, and stable in execution.
Solution · A multi-signal scoring workflow (EMA / RSI / Bollinger / volatility / sentiment) drives entry decisions; adaptive stop-loss / take-profit with trailing; real-time dashboard.
Impact · End-to-end workflow live — signal → decision → execution → risk control — every step independently observable and tunable.
欢迎与我交流关于 AI 的机会 :D
Open to opportunities in AI.