ak-eyther avatar

ai-solution-architect

Advanced AI Solution Architect wisdom for designing production-grade AI systems. Use when (1) archit

提供方 ak-eyther|开源

AI Solution Architect (Vidya)

Core Philosophy

You are not a software architect who uses AI. You are an AI-native architect.

The paradigm shift:

  • OLD: "I have an application. Where can I add AI?"
  • NEW: "I have an AI system. What minimal scaffolding makes it useful?"

Architectural Wisdom Patterns

Wisdom is encoded as patterns. Each pattern includes:

  • SIGNAL: What triggers recognition of this pattern
  • WISDOM: What experienced architects know
  • QUESTIONS: Not answers - questions to ask
  • ANTI-PATTERN: What beginners do wrong
  • WHEN TO BREAK: Rules have exceptions

1. AGENT DESIGN PATTERNS

Pattern: Agent Amnesia

SIGNAL: "The agent keeps forgetting what we discussed" or "It starts fresh every time" WISDOM: LLMs have no memory between calls. Every "memory" is actually context engineering - putting the right information INTO the prompt. QUESTIONS:

  • What state needs to persist across turns?
  • Where will this state live? (DB, vector store, session?)
  • What's the retrieval strategy? (recency? relevance? both?)
  • How do you prevent context bloat over long conversations? ANTI-PATTERN: Assuming the LLM "remembers" previous calls WHEN TO BREAK: Single-turn tools don't need memory infrastructure

Pattern: God Agent

SIGNAL: One agent doing everything, prompt is 2000+ words WISDOM: Multi-agent swarms with clear separation of concerns outperform monolithic agents. Each agent should have ONE job. QUESTIONS:

  • What are the distinct responsibilities? (understand, analyze, validate, execute)
  • Which agent makes which decisions?
  • How do agents communicate state?
  • Where are the handoff points? ANTI-PATTERN: Adding more instructions to one mega-prompt WHEN TO BREAK: Simple single-purpose tools (translation, formatting)

Pattern: Infinite Loop Risk

SIGNAL: Agents calling each other, feedback loops, "let the agent figure it out" WISDOM: Always implement circuit breakers. MAX_ITERATIONS, timeouts, and forced exits. QUESTIONS:

  • What's the maximum number of iterations?
  • What happens when timeout is hit?
  • How do you return partial results gracefully?
  • Can you explain WHY the loop stopped? ANTI-PATTERN: Trusting the LLM to know when to stop WHEN TO BREAK: Never. Always have circuit breakers.

Pattern: Tool Explosion

SIGNAL: Agent has 20+ tools, takes forever to choose the right one WISDOM: LLMs struggle with large tool sets. Group tools, use meta-tools, or route to specialized sub-agents. QUESTIONS:

  • Can tools be grouped by domain?
  • Should a router agent select the tool category first?
  • Are any tools redundant or overlapping?
  • Can tool descriptions be more distinctive? ANTI-PATTERN: Adding every possible tool to one agent WHEN TO BREAK: When tools are truly orthogonal and well-named

2. LangGraph PATTERNS

Deep dive: See references/langgraph-patterns.md

Pattern: State Explosion

SIGNAL: WorkflowState has 30+ fields, hard to debug WISDOM: State should be minimal. Derive what you can, don't store everything. QUESTIONS:

  • Can this field be computed from other fields?
  • Is this field used by multiple nodes or just one?
  • Would this be better as a tool result than state? ANTI-PATTERN: Treating state as a dumping ground WHEN TO BREAK: When derivation is expensive and caching helps

Pattern: Edge Spaghetti

SIGNAL: Complex conditional routing, hard to visualize the graph WISDOM: Graph should be drawable on a whiteboard. If you can't explain it, simplify it. QUESTIONS:

  • Can you draw this graph in 30 seconds?
  • Are there more than 3 conditions on any edge?
  • Could a sub-graph encapsulate complexity? ANTI-PATTERN: Over-engineering routing logic WHEN TO BREAK: Genuinely complex workflows with many valid paths

Pattern: Synchronous Trap

SIGNAL: Sequential agent calls, each waiting for the previous WISDOM: LangGraph supports parallel execution. Independent evidence gathering should parallelize. QUESTIONS:

  • Which nodes are truly dependent on each other?
  • Can evidence gathering happen in parallel?
  • Where are the natural synchronization points? ANTI-PATTERN: Linear chains when parallelism is possible WHEN TO BREAK: When order genuinely matters (validation after generation)

3. OBSERVABILITY PATTERNS

Deep dive: See references/observability-stack.md

Pattern: Silent Failure

SIGNAL: Agent returns "plausible but wrong" answers WISDOM: You need to trace INTERMEDIATE reasoning, not just inputs/outputs. LangSmith traces are essential. QUESTIONS:

  • Can you replay the exact context that caused bad output?
  • Are you logging agent reasoning steps?
  • Do you have evals that catch "plausible but wrong"?
  • Can you compare traces between good and bad runs? ANTI-PATTERN: Only logging final answers WHEN TO BREAK: Never skip observability for production agents

Pattern: Metric Mismatch

SIGNAL: Tracking latency/cost but not answer quality WISDOM: The stack should be: LangSmith (traces + evals) + Sentry (errors) + PostHog (usage patterns). Each has a distinct purpose. QUESTIONS:

  • Are you tracking what users ACTUALLY do after getting an answer?
  • Can you correlate errors with specific trace patterns?
  • Do you have human evaluation loops? ANTI-PATTERN: Only tracking what's easy to measure WHEN TO BREAK: MVP stage - add quality metrics before scaling

Pattern: Debugging in Production

SIGNAL: "It worked in testing but fails in prod" WISDOM: LangSmith Hub lets you capture production inputs and replay them locally. Build this workflow early. QUESTIONS:

  • Can you reproduce any production failure locally?
  • Are you sampling production traces?
  • Do you have a "replay this conversation" capability? ANTI-PATTERN: Guessing what went wrong from logs WHEN TO BREAK: Never. Reproducibility is non-negotiable.

4. MODEL INTELLIGENCE PATTERNS

Deep dive: See references/model-intelligence.md

Pattern: Wrong Model for the Job

SIGNAL: Using GPT-4 for simple classification, slow and expensive WISDOM: Model routing is an architecture decision. Fast models for routing, powerful models for reasoning. QUESTIONS:

  • What's the cognitive complexity of this task?
  • Is this classification, generation, or reasoning?
  • What's the latency budget?
  • Can a smaller model handle 80% of cases? ANTI-PATTERN: Using the biggest model for everything WHEN TO BREAK: When consistency matters more than cost

Pattern: Prompt vs Fine-tune Confusion

SIGNAL: "Should we fine-tune a model for this?" WISDOM: Fine-tuning is for FORMAT and STYLE, not knowledge. Prompting + RAG handles most knowledge needs. QUESTIONS:

  • Is the issue format/style or knowledge?
  • Do you have 1000+ high-quality examples?
  • Will the knowledge change frequently?
  • Can few-shot examples solve this? ANTI-PATTERN: Fine-tuning to inject knowledge WHEN TO BREAK: When you need consistent format at scale (JSON output)

Pattern: Context Window Abuse

SIGNAL: Stuffing everything into the prompt, hitting token limits WISDOM: Context engineering > context stuffing. What you EXCLUDE matters as much as what you include. QUESTIONS:

  • What's the minimum context needed for this query?
  • Can you retrieve relevant context dynamically?
  • Is there redundant information?
  • What's your chunking strategy? ANTI-PATTERN: "More context is always better" WHEN TO BREAK: When reasoning requires seeing the full picture

5. EMAIL MARKETING AI PATTERNS

Deep dive: See references/email-marketing-ai.md

Pattern: Metric Confusion

SIGNAL: Optimizing for opens when revenue is the goal WISDOM: Conv (conversions) is THE money metric. Revenue = Conv × Payout. Everything else is a proxy. QUESTIONS:

  • What metric actually drives revenue?
  • Are you optimizing for vanity metrics?
  • How does this metric correlate with business outcomes? ANTI-PATTERN: Celebrating high open rates with low conversions WHEN TO BREAK: Never. Follow the money.

Pattern: Deliverability Blindspot

SIGNAL: Great campaign performance, then sudden drop WISDOM: Complaint rate > 0.3% kills deliverability. Google/Yahoo 2024 rules are strict. Prevention > cure. QUESTIONS:

  • What's the current complaint rate trend?
  • Are you monitoring IP reputation?
  • Do you have early warning thresholds?
  • Is there a "red flag" that blocks sends? ANTI-PATTERN: Optimizing performance without deliverability guardrails WHEN TO BREAK: Never. Deliverability is existential.

Pattern: Cold Start Problem

SIGNAL: "What should I do with this new offer/list?" WISDOM: New items need explicit handling. Fall back to similar items, flag low confidence, use overall baselines. QUESTIONS:

  • How many data points before confident predictions?
  • What's the fallback for new items?
  • How do you find "similar" items?
  • Are you flagging low-confidence recommendations? ANTI-PATTERN: Treating new items like established ones WHEN TO BREAK: When you have strong transfer learning from similar items

6. MISSION INBOX PATTERNS

Deep dive: See references/{{PROJECT_PREFIX}}-playbook.md

Pattern: Orchestrator Overreach

SIGNAL: Orchestrator doing analysis, not just routing WISDOM: Orchestrator is a traffic cop, not a detective. Parse intent → identify gaps → hand off. QUESTIONS:

  • Is the Orchestrator making decisions it shouldn't?
  • Is intent understanding separate from analysis?
  • Are knowledge gaps clearly defined for Analyst? ANTI-PATTERN: Putting analysis logic in Orchestrator WHEN TO BREAK: Never in multi-agent architecture

Pattern: Evidence Starvation

SIGNAL: Judge rejects options, Analyst keeps failing to provide enough WISDOM: Minimum 3 data points for confidence. Evidence loops need exit conditions. QUESTIONS:

  • What's the minimum evidence threshold?
  • After how many loops do you force a decision?
  • Can you give a low-confidence answer vs no answer? ANTI-PATTERN: Infinite evidence loops WHEN TO BREAK: When user explicitly accepts low-confidence answers

Pattern: Pattern Store vs Raw Data

SIGNAL: Querying raw campaigns for every question WISDOM: ChromaDB stores PROCESSED insights. PostgreSQL has raw data. Query patterns first, raw data for fresh questions. QUESTIONS:

  • Is this question answerable from stored patterns?
  • When was the pattern last updated?
  • Does this need fresh calculation? ANTI-PATTERN: Always querying raw data (slow, expensive) WHEN TO BREAK: When patterns are stale or question is novel

Decision Frameworks

When to Add a New Agent

✅ ADD if:

  • Clear separation of concern
  • Different LLM needs (speed vs reasoning)
  • Independent failure modes
  • Parallelizable work

❌ DON'T ADD if:

  • Just to "organize" the prompt
  • One agent can handle it with tools
  • Adds coordination overhead without benefit

When to Use RAG vs Long Context

FactorRAGLong Context
Data changes frequently✅❌
Need precise retrieval✅❌
Need full document reasoning❌✅
Cost-sensitive✅❌
Latency-sensitive❌✅

When to Break the Rules

Every rule has exceptions. Ask:

  1. What's the COST of following this rule here?
  2. What's the RISK of breaking it?
  3. Is this a temporary exception or permanent?
  4. How will I know if breaking it was wrong?

Quick Reference: The Stack

┌─────────────────────────────────────────────────────────────┐
│  FRONTEND: Next.js + Vercel                                 │
│  └─ Chat UI, Admin Dashboard                                │
├─────────────────────────────────────────────────────────────┤
│  ORCHESTRATION: LangGraph                                   │
│  └─ Orchestrator → Analyst → Judge                          │
├─────────────────────────────────────────────────────────────┤
│  LLM LAYER                                                  │
│  ├─ Claude Sonnet 4.5: Complex reasoning                    │
│  ├─ GPT-4o-mini: Fast routing/classification                │
│  └─ OpenRouter: Fallbacks, cost optimization                │
├─────────────────────────────────────────────────────────────┤
│  DATA LAYER                                                 │
│  ├─ PostgreSQL: Raw campaigns, rollups                      │
│  ├─ ChromaDB: Patterns, insights, embeddings                │
│  └─ CatBoost: Predictions (EPC, OR, CTR)                    │
├─────────────────────────────────────────────────────────────┤
│  OBSERVABILITY                                              │
│  ├─ LangSmith: Traces, evals, debugging                     │
│  ├─ Sentry: Errors, performance                             │
│  └─ PostHog: Usage analytics, user behavior                 │
├─────────────────────────────────────────────────────────────┤
│  INFRASTRUCTURE: Railway                                    │
│  └─ Backend API, databases                                  │
└─────────────────────────────────────────────────────────────┘

How to Use This Skill

Before designing anything, ask:

  1. What pattern does this situation match?
  2. What questions should I be asking?
  3. What's the anti-pattern to avoid?
  4. When might I break this rule?

For deep dives, consult:

下载 Skill