ai-solution-architecture
Design pre-sales technical architectures for AI/ML solutions targeting B2B enterprise clients. Use when scoping AI implementations, designing reference architectures, mapping integration patterns, est
SKILL.mdplugins/loxihart-ai/skills/ai-solution-architecture/SKILL.md
GitHubAI Solution Architecture
Purpose
Design comprehensive, production-ready AI/ML solution architectures for enterprise B2B clients during pre-sales cycles. Translates business requirements into technical designs with infrastructure sizing, integration patterns, security considerations, and cost modeling.
Trigger Patterns
- "Design an AI solution for..."
- "What architecture do we need for..."
- "Scope this AI implementation"
- "Technical proposal for [AI use case]"
- "Integration architecture for AI"
- "Infrastructure requirements for ML"
- Client mentions: solution design, technical scoping, reference architecture, AI implementation plan
Input Requirements
required:
- business_objective: "What problem is AI solving?"
- scale_parameters: "Users, requests/day, data volume"
- existing_systems: "Current tech stack, APIs, databases"
optional:
- compliance_requirements: "GDPR, HIPAA, [COMPLIANCE FRAMEWORK], etc."
- budget_constraints: "Infrastructure budget, timeline"
- team_capabilities: "In-house AI expertise level"
Process
1. Requirements Discovery
Business Context Analysis
- Problem statement: What manual process is being automated?
- Success metrics: How will ROI be measured?
- User personas: Who interacts with the AI?
- Scale requirements: Peak load, data growth, latency SLOs
Technical Constraints Mapping
- Existing infrastructure: Cloud provider (AWS/Azure/GCP), on-prem
- Integration points: CRM, ERP, databases, APIs
- Security posture: Authentication, authorization, data encryption
- Compliance mandates: Industry regulations, data residency
2. Architecture Design
Component stack options — LLM layer, application framework, data layer, and infrastructure — with selection criteria. See references/stack-reference.md for the full stack comparison tables.
3. Integration Mapping
Authentication, authorization, and API integration pattern options. See references/stack-reference.md for the full integration pattern tables.
4. Infrastructure Sizing
Request Volume Modeling
See references/sizing-calculator.md for detailed methodology.
Quick Estimates
Small / medium / large deployment tiers with monthly cost and hosting breakdowns. See references/stack-reference.md for the full pricing tier and hosting tables.
5. Security Architecture
Threat Model
threats:
prompt_injection:
mitigation: Input sanitization, content filtering, system prompt isolation
data_exfiltration:
mitigation: Network policies, egress filtering, audit logging
unauthorized_access:
mitigation: JWT validation, rate limiting, IP allowlisting
model_poisoning:
mitigation: Frozen base models, validation of fine-tuning data
Compliance Mapping
See references/compliance-patterns.md for detailed frameworks.
Output Format
Architecture Diagram (Mermaid)
graph TB
User[End Users] -->|HTTPS| CDN[Cloudflare CDN]
CDN --> LB[Load Balancer]
LB --> API1[API Server 1]
LB --> API2[API Server 2]
API1 --> Auth[Auth Service<br/>Clerk]
API2 --> Auth
API1 --> LLM[LLM Gateway<br/>LangChain]
API2 --> LLM
LLM --> OpenAI[OpenAI API<br/>GPT-4]
LLM --> VectorDB[Pinecone<br/>Vector Store]
API1 --> DB[(PostgreSQL<br/>Supabase)]
API2 --> DB
API1 --> Cache[(Redis<br/>Upstash)]
API2 --> Cache
DB --> Backup[S3 Backups]
API1 --> Monitor[PostHog<br/>Analytics]
API2 --> Monitor
API1 --> Errors[Sentry<br/>Error Tracking]
API2 --> Errors
Technical Specification Document
# AI Solution Architecture: [Client Name] - [Use Case]
**Date:** 2025-01-22
**Version:** 1.0
**Prepared by:** Loxihart Inc. Sales Engineering
---
## Executive Summary
**Business Objective:** [1-2 sentences]
**Proposed Solution:** [Technology stack summary]
**Estimated Timeline:** [Weeks to production]
**Total Cost:** $[monthly recurring] + $[one-time implementation]
---
## Architecture Overview
### Component Stack
- **LLM Provider:** [OpenAI GPT-4 / Claude 3.5 Sonnet / Self-hosted Llama 3.1]
- **Application Framework:** [FastAPI / Next.js / Custom]
- **Vector Database:** [Pinecone / Weaviate / pgvector]
- **Relational Database:** [PostgreSQL / MySQL]
- **Authentication:** [Clerk / Auth0 / Custom JWT]
- **Hosting:** [Vercel / AWS / GCP]
### Integration Points
1. **[System A]:** [Integration method, data flow direction]
2. **[System B]:** [Integration method, data flow direction]
---
## Infrastructure Sizing
**Expected Load:**
- Daily Active Users: [number]
- Requests per Day: [number]
- Peak QPS: [number]
- Data Storage: [GB/TB]
**Provisioned Resources:**
- Compute: [Instance types, count]
- Database: [Size, IOPS]
- Vector DB: [Dimensions, vectors count]
---
## Security & Compliance
**Data Classification:** [Public / Internal / Confidential / Restricted]
**Compliance Requirements:**
- [ ] GDPR (EU data residency, right to deletion)
- [ ] HIPAA (PHI encryption, audit logging)
- [ ] [COMPLIANCE FRAMEWORK] (access controls, monitoring)
**Security Controls:**
- Authentication: [Multi-factor, SSO]
- Authorization: [RBAC, RLS policies]
- Encryption: [At-rest AES-256, In-transit TLS 1.3]
- Audit Logging: [All API requests logged, 90-day retention]
---
## Cost Breakdown
| Component | Monthly Cost | Notes |
|-----------|--------------|-------|
| Hosting (Vercel Pro) | $20 | Scales to 100GB bandwidth |
| Database (Supabase Pro) | $25 | 8GB RAM, 100GB storage |
| Vector DB (Pinecone) | $70 | 1 pod, 1M vectors |
| LLM API (OpenAI) | $500-2,000 | Variable by usage |
| Observability (PostHog) | $0-50 | Free tier, then usage-based |
| **Total** | **$615-2,165/month** | |
**One-Time Costs:**
- Implementation (Loxihart Services): $[amount]
- Data migration: $[amount]
- Training: $[amount]
---
## Deployment Phases
### Phase 1: Foundation (Weeks 1-2)
- Setup infrastructure (VPC, databases, hosting)
- Configure authentication (SSO, user management)
- Deploy base API (health checks, monitoring)
### Phase 2: Core AI Features (Weeks 3-4)
- LLM integration (prompts, context management)
- Vector database setup (embeddings, search)
- API endpoints (chat, document processing)
### Phase 3: Integration (Weeks 5-6)
- Connect to [Client System A]
- Connect to [Client System B]
- Data pipeline setup (ETL, synchronization)
### Phase 4: Testing & Launch (Weeks 7-8)
- Load testing (target: 1,000 concurrent users)
- Security audit (penetration testing)
- Production deployment (blue/green rollout)
---
## Risks & Mitigations
| Risk | Likelihood | Impact | Mitigation |
|------|------------|--------|------------|
| LLM API rate limits | Medium | High | Implement retry logic, fallback models |
| Data migration delays | High | Medium | Parallel track, early testing |
| Integration complexity | Medium | Medium | Dedicated integration engineer, buffer time |
| Cost overruns | Medium | Medium | Usage monitoring, alerts at 80% budget |
---
## Success Criteria
- [ ] P95 latency < 2 seconds
- [ ] 99.9% uptime (max 43 minutes downtime/month)
- [ ] < 0.5% error rate
- [ ] User satisfaction > 4.5/5
- [ ] Processes [target number] requests/day
---
## Next Steps
1. **Client:** Review and approve architecture
2. **Loxihart:** Finalize SOW (Statement of Work)
3. **Client:** Provision accounts (cloud, LLM providers)
4. **Loxihart:** Begin Phase 1 implementation
**Questions? Contact:** [EMAIL]
Validation Gates
architecture_review:
- [ ] All integration points documented
- [ ] Security controls mapped to threats
- [ ] Infrastructure sized for peak + 30% buffer
- [ ] Costs broken down by component
- [ ] Compliance requirements addressed
technical_review:
- [ ] No single points of failure
- [ ] Observability hooks in all components
- [ ] Graceful degradation patterns defined
- [ ] Rollback strategy documented
business_review:
- [ ] ROI justification included
- [ ] Timeline realistic (client validation)
- [ ] Risks quantified (likelihood × impact)
- [ ] Success metrics measurable
Error Handling
| Error | Cause | Fix |
|---|---|---|
| "Insufficient scale data" | Client doesn't know expected load | Provide 3 sizing tiers (S/M/L), let client choose |
| "Integration unclear" | Legacy system documentation missing | Schedule technical discovery call, request API docs |
| "Budget misalignment" | Proposed solution exceeds budget | Offer phased approach, start with MVP |
| "Compliance unknown" | Client unsure of regulatory requirements | Provide compliance questionnaire, reference common frameworks |
Cross-References
- See
roi-calculator-buildingfor cost/benefit analysis - See
competitive-battle-cardsfor positioning vs alternatives - See
poc-scopingfor proof-of-concept design - See
technical-discoveryfor requirements gathering templates
Version: 1.1.0
Last Updated: 2025-01-22
Maintainer: Loxihart Inc. Sales Engineering
Support: For questions, reference /references/ or contact [EMAIL]
Reference Files
references/stack-reference.md— full stack comparison tables (LLM, application framework, data layer, infrastructure), integration pattern options, and deployment pricing/hosting tiers (moved from sections 2–4).references/sizing-calculator.md— detailed infrastructure sizing methodology.references/compliance-patterns.md— detailed compliance frameworks.