AI Solutions Architect Portfolio

End-to-end AWS Bedrock cost optimization: 4 strategies benchmarked, real CloudWatch monitoring, and a live AI Daily Digest pipeline — all in one monorepo.

Architecture

graph TB
    subgraph Client["Client / Application"]
        REQ["User Request"]
    end

    subgraph Optimization["Optimization Layer"]
        COMP["Prompt Compression<br/>RelevanceFilter · LLMLingua-2"]
        CACHE["Prompt Caching<br/>System prompt reuse · 90% read discount"]
        ROUTE["Model Routing<br/>Haiku (classify) → Sonnet (generate)"]
        BATCH["Batch Processing<br/>50% off · async jobs"]
    end

    subgraph Bedrock["Amazon Bedrock"]
        HAIKU["Claude Haiku 4.5<br/>$1 / $5 per 1M tokens"]
        SONNET["Claude Sonnet 4<br/>$3 / $15 per 1M tokens"]
    end

    subgraph Monitoring["Observability"]
        MET["Metrics Collector<br/>put_metric_data"]
        CW["CloudWatch<br/>BedrockCost Namespace"]
        DASH["Dashboard<br/>Cost · Tokens · Calls"]
        ALARM["Alarm ($1/day)<br/>→ SNS → Email"]
        CE["Cost Explorer API<br/>7-day trend"]
    end

    REQ --> COMP & CACHE & ROUTE & BATCH
    COMP --> SONNET
    CACHE --> SONNET
    ROUTE --> HAIKU & SONNET
    BATCH --> HAIKU & SONNET

    HAIKU & SONNET --> MET
    MET --> CW
    CW --> DASH
    CW --> ALARM
    CE -.-> CW

Loading

Key Results

Benchmarked on Claude Sonnet 4 with LongBench dataset (2K–10K word contexts):

Strategy	Technique	Cost Reduction	Quality Impact
Prompt Compression	RelevanceFilter (TF-IDF + Jaccard)	76.3%	No degradation (4/5 judge pass)
Prompt Compression	LLMLingua-2 (local BERT)	34.4%	Minor (3/5 judge pass)
Model Routing	Haiku for scoring, Sonnet for generation	~25x cheaper on filtering step	Equivalent for classification
Prompt Caching	System prompt cache reads	20.6%	No impact
Batch Processing	Async batch API	50.0%	No impact

Full benchmark: testing/BENCHMARK_RESULTS_SUMMARY.md · Model selection guide: docs/model-selection-matrix.md

Prerequisites

• Python ≥ 3.13
• uv — package manager
• Node.js — required for AWS CDK CLI
• AWS CLI configured with credentials (aws sts get-caller-identity should work)
• AWS IAM permissions:

  bedrock:InvokeModel
  bedrock:InvokeModelWithResponseStream
  cloudwatch:PutMetricData
  sns:Publish
  ce:GetCostAndUsage
  

• Bedrock model access enabled for: Claude Haiku 4.5, Claude Sonnet 4

Quick Start

# Clone and setup
git clone https://github.com/yepengfan/ai-sa-portfolio.git
cd ai-sa-portfolio/systems/s1-cost

# Install dependencies
uv sync --extra dev

# Run the AI Daily Digest (fetches RSS, scores, summarizes, writes to Obsidian)
uv run python -m digest

# Run with dry-run (no AWS, stdout only)
uv run python -m digest --stdout --no-metrics --top-n 3

# Deploy CloudWatch infrastructure (Dashboard + Alarm + SNS)
cd infra && npx aws-cdk@latest deploy --app "uv run python3 app.py"

# Run tests
uv run pytest digest/tests/ -v         # 21 digest tests
cd infra && uv run pytest tests/ -v    # 10 CDK tests

Project Structure

ai-sa-portfolio/
├── systems/
│   └── s1-cost/                        # System #1 — Cost Optimization
│       ├── digest/                     # AI Daily Digest pipeline
│       │   ├── main.py                 # Pipeline orchestration + CLI
│       │   ├── feeds.py                # 92 Karpathy-curated RSS feeds
│       │   ├── sources/rss.py          # Async RSS fetcher
│       │   ├── dedup.py                # Jaccard dedup
│       │   ├── scoring.py              # Haiku batch scoring
│       │   ├── summarizer.py           # Sonnet bilingual summary
│       │   └── report.py              # Obsidian markdown output
│       ├── infra/                      # CDK infrastructure
│       │   ├── stacks/monitoring_stack.py
│       │   └── tests/
│       ├── utils/                      # Shared utilities
│       │   ├── bedrock.py              # Converse API wrapper
│       │   ├── metrics.py              # CloudWatch metrics
│       │   └── config.py               # Model IDs + pricing
│       ├── benchmark/                  # LongBench experiments
│       ├── strategies/                 # Compression implementations
│       ├── docs/                       # Model selection matrix
│       └── testing/                    # Benchmark results
├── portfolio/                          # Portfolio projects (planned)
└── docs/                               # Design documents

Systems Overview

System #1: Cost Optimization (s1-cost/)

Bedrock Optimization Strategies — 4 cost reduction strategies benchmarked and deployed

• Prompt Compression: 5 techniques (RelevanceFilter, LLMLingua-2, Semantic Summarizer, Structure Optimizer, Manual Refiner)
• Prompt Caching: System prompt reuse with 90% read discount
• Model Routing: Haiku-as-filter, Sonnet-as-generator pattern
• Batch Processing: 50% off via async batch API
• AI Daily Digest: Live pipeline — 92 RSS feeds → Haiku scoring → Sonnet bilingual summary → Obsidian
• CloudWatch Monitoring: CDK-deployed Dashboard, Alarm ($1/day), SNS email alerts
• Status: ✅ Complete

System #2–4 (Planned)

System	Focus	Status
s2-pipeline	ML/AI Pipeline & Orchestration	Planned
s3-ha	High Availability & Fault Tolerance	Planned
s4-observability	Observability Platform	Planned

Blog

AI Inference Eats 80% of Your Budget — I Benchmarked 4 Strategies to Cut It

Published on Medium. Covers all 4 optimization strategies with real benchmark data from Claude Sonnet 4 on Amazon Bedrock.

Roadmap

• ✅ Phase 1: s1-cost — Bedrock cost optimization (compression, caching, routing, batch)
• ✅ Phase 1: AI Daily Digest — Live Bedrock pipeline with CloudWatch monitoring
• ✅ Phase 1: CloudWatch infrastructure — CDK Dashboard + Alarm + SNS
• Next: s2-pipeline — ML orchestration and data processing
• Future: Complete all 4 systems + 3 portfolio projects