Mini-me

Project Type

App Product Design

Design Stage

Startup

Deliverables

Motion design system, UI Kit, Applied Motion Design, motional Support Strategy, Product Design

1. Defining the Agent’s Role

Generative models are no longer the bottleneck. Human expression is. Users often know exactly what they want, but struggle to translate that intent into explicit instructions AI can understand. AIONE exists to bridge this gap—transforming human intent into structured, interpretable instructions while keeping people in control of the process.

Mini-me vs chatbot

AIONE multi-layer workflow:

2. Designing the Agent System

Generative models are no longer the bottleneck. Human expression is. Users often know exactly what they want, but struggle to translate that intent into explicit instructions AI can understand. AIONE exists to bridge this gap—transforming human intent into structured, interpretable instructions while keeping people in control of the process.

Existing Creative Workflow: While AI has become part of the creative process, users still face significant uncertainty because prompt quality heavily influences the final result.

AIONE multi-layer workflow:

Existing Solutions on the Market: Expand user keywords into longer, more elaborate prompts, but often fail to capture the user's actual intent.

Instead of simply expanding keywords, AIONE helps users clarify, structure, and refine their intent. Guided recommendations, visual references, version control, and human-in-the-loop collaboration ensure that prompts remain aligned with user goals.

Zoom In 1：Designing the Knowledge Architecture

这部分不应只写“整理了数据”，而要展示你如何把个人经历转化成可供 agent 使用的知识系统。

核心问题

How do you turn an entire professional history into knowledge an AI agent can reliably understand and retrieve?

可以展示的内容

  • 将知识拆成 profile、experience、capabilities、FAQ、conversation guidance 和 projects

  • 为什么普通 portfolio 文本不能直接拿来做 retrieval

  • 每个 chunk 必须：

    • self-contained

    • one primary idea

    • retrieval-friendly

    • traceable to source

    • stable over time

  • Markdown 与 structured JSON 的不同 ingestion

  • 为 identity 信息增加独立的 profile.json ingestion

  • 从 95 个 chunks 增加到 99 个 chunks

  • 统一 schema、question ID、category、keywords、related questions

  • 如何处理项目事实、职业身份和抽象能力之间的交叉关系

最适合的视觉

做一个从原始资料到 agent knowledge 的 transformation diagram：

Resume / Portfolio / Project Docs / FAQ
                    ↓
        Knowledge Taxonomy
                    ↓
     Self-contained Knowledge Chunks
                    ↓
     Metadata + Source + Relationships

旁边放一个“Bad Chunk vs Good Chunk”对比，会非常有说服力。

这一部分体现的能力

  • Information Architecture

  • Knowledge Design

  • Systems Thinking

  • Content Governance

  • AI-native Product Design

3. Designing the Knowledge Architecture

Generative models are no longer the bottleneck. Human expression is. Users often know exactly what they want, but struggle to translate that intent into explicit instructions AI can understand. AIONE exists to bridge this gap—transforming human intent into structured, interpretable instructions while keeping people in control of the process.

Existing Creative Workflow: While AI has become part of the creative process, users still face significant uncertainty because prompt quality heavily influences the final result.

AIONE multi-layer workflow:

Existing Solutions on the Market: Expand user keywords into longer, more elaborate prompts, but often fail to capture the user's actual intent.

Instead of simply expanding keywords, AIONE helps users clarify, structure, and refine their intent. Guided recommendations, visual references, version control, and human-in-the-loop collaboration ensure that prompts remain aligned with user goals.

Zoom In 2：Making Retrieval Reliable

这是 Mini-Me 最核心的技术深度点。

核心问题

How do you make sure the agent retrieves the right evidence instead of merely finding semantically similar text?

可以展示的内容

完整流程：

User Question
→ Embedding
→ Supabase pgvector Search
→ Similarity Scoring
→ Top Relevant Chunks
→ Context Assembly
→ Grounded Response

具体证据可以保留：

  • OpenAI text-embedding-3-small

  • 1,536-dimensional embeddings

  • Supabase + pgvector

  • Top-K 5

  • similarity threshold 0.50

  • 99 knowledge chunks

  • representative retrieval evaluation

但重点不是罗列参数，而是解释你如何发现 retrieval 失败：

  • “Tell me about Jasper” 最初没有 profile chunks

  • Mini-Me role 信息存在，但 chunk 不独立，导致没有被召回

  • architecture background 和 trustworthy AI project evidence 检索不足

  • 宽泛关键词可能召回语义接近但 intent 错误的内容

  • 单次成功不能证明系统有效，必须建立 evaluation set

然后展示你如何修复：

  • 增加缺失知识

  • 重写 chunk context

  • 调整 chunk boundaries

  • 增加 structured profile ingestion

  • 分离 identity handling 与普通 retrieval

  • 检查 similarity scores 和 returned chunks

  • 重建 embeddings 与 database

最适合的视觉

不要只画普通 RAG 流程图。更有价值的是做一个 retrieval debugging example：

Question:
“Tell me about Jasper.”

Before
→ No profile chunk
→ Project-related result
→ Incomplete answer

Diagnosis
→ Missing structured identity knowledge

After
→ Profile ingestion
→ Correct profile chunks retrieved
→ Grounded professional introduction

它能证明你真正调试过 RAG，而不是仅仅调用了 API。

这一部分体现的能力

  • RAG Product Design

  • Retrieval Evaluation

  • AI Engineering Collaboration

  • Technical Fluency

  • Evidence-based Iteration

4. Designing Trust and Control

Generative models are no longer the bottleneck. Human expression is. Users often know exactly what they want, but struggle to translate that intent into explicit instructions AI can understand. AIONE exists to bridge this gap—transforming human intent into structured, interpretable instructions while keeping people in control of the process.

Existing Creative Workflow: While AI has become part of the creative process, users still face significant uncertainty because prompt quality heavily influences the final result.

AIONE multi-layer workflow:

Existing Solutions on the Market: Expand user keywords into longer, more elaborate prompts, but often fail to capture the user's actual intent.

Instead of simply expanding keywords, AIONE helps users clarify, structure, and refine their intent. Guided recommendations, visual references, version control, and human-in-the-loop collaboration ensure that prompts remain aligned with user goals.

5. From Design to Working Product

AIONE's goal is not to replace human decision-making, but to help users express, refine, and preserve their intent throughout the creative process. At every stage, users can accept, modify, ignore, or replace AI recommendations. Strong intent signals are preserved, while weak or missing dimensions become opportunities for guided exploration rather than automated decisions. Users remain responsible for selecting directions, evaluating alternatives, and determining which ideas best reflect their goals. Visual vibe matching further supports this process by helping users evaluate concepts through imagery rather than language alone. The system evaluates, guides, and explains, but creative ownership always remains with the user.

Human-in-the-Loop Creative Process Workflow

AIONE keeps humans in control throughout the entire creative process. From intent steering and vibe exploration to manual refinement and version comparison, users actively shape every decision while AI provides guidance, context, and alternatives. The system amplifies human judgment rather than replacing it.