Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
How we built an AI code reviewer with serverles...
Search
Yan Cui
February 12, 2025
Technology
0
110
How we built an AI code reviewer with serverless and Bedrock
Slides for my talk at the Serverless London meetup on 12-Feb-2025
Yan Cui
February 12, 2025
Tweet
Share
More Decks by Yan Cui
See All by Yan Cui
Money-saving tips for the frugal serverless developer (AWS Community Summit)
theburningmonk
1
190
Money-saving tips for the frugal serverless developer
theburningmonk
1
770
Why the fuzz about serverless (with CompassDigital)
theburningmonk
0
110
Money-saving tips for the frugal serverless developer
theburningmonk
0
130
Efficient patterns for serverless development (AWS Summit London)
theburningmonk
0
150
7 ways to solve Lambda cold starts
theburningmonk
0
65
Saving Money on Serverless: Common Mistakes and How to Avoid Them
theburningmonk
0
58
3 Ways to Improve Serverless Performance
theburningmonk
0
46
Smart and efficient ways to test serverless architectures
theburningmonk
1
290
Other Decks in Technology
See All in Technology
roppongirb_20250911
igaiga
1
250
「何となくテストする」を卒業するためにプロダクトが動く仕組みを理解しよう
kawabeaver
0
430
OCI Oracle Database Services新機能アップデート(2025/06-2025/08)
oracle4engineer
PRO
0
180
フルカイテン株式会社 エンジニア向け採用資料
fullkaiten
0
8.8k
会社紹介資料 / Sansan Company Profile
sansan33
PRO
6
380k
Evolución del razonamiento matemático de GPT-4.1 a GPT-5 - Data Aventura Summit 2025 & VSCode DevDays
lauchacarro
0
210
S3アクセス制御の設計ポイント
tommy0124
3
200
Webアプリケーションにオブザーバビリティを実装するRust入門ガイド
nwiizo
7
890
機械学習を扱うプラットフォーム開発と運用事例
lycorptech_jp
PRO
0
660
20250910_障害注入から効率的復旧へ_カオスエンジニアリング_生成AIで考えるAWS障害対応.pdf
sh_fk2
3
280
Claude Code でアプリ開発をオートパイロットにするためのTips集 Zennの場合 / Claude Code Tips in Zenn
wadayusuke
5
1.6k
「その開発、認知負荷高すぎませんか?」Platform Engineeringで始める開発者体験カイゼン術
sansantech
PRO
2
720
Featured
See All Featured
GitHub's CSS Performance
jonrohan
1032
460k
Being A Developer After 40
akosma
90
590k
Practical Orchestrator
shlominoach
190
11k
How to Create Impact in a Changing Tech Landscape [PerfNow 2023]
tammyeverts
53
3k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
667
120k
[Rails World 2023 - Day 1 Closing Keynote] - The Magic of Rails
eileencodes
36
2.5k
Context Engineering - Making Every Token Count
addyosmani
3
60
Measuring & Analyzing Core Web Vitals
bluesmoon
9
580
Distributed Sagas: A Protocol for Coordinating Microservices
caitiem20
333
22k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
jQuery: Nuts, Bolts and Bling
dougneiner
64
7.9k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
358
30k
Transcript
How we built an AI Code Reviewer with Serverless and
Bedrock
Yan Cui http://theburningmonk.com @theburningmonk AWS user since 2010
Yan Cui http://theburningmonk.com @theburningmonk running serverless in production since 2016
Developer Advocate @ Yan Cui http://theburningmonk.com @theburningmonk
Yan Cui http://theburningmonk.com @theburningmonk independent consultant
None
evolua.io Demo
Architecture
API Gateway EventBridge Webhook
API Gateway DynamoDB Bedrock EventBridge Webhook
API Gateway DynamoDB Bedrock EventBridge Webhook
API Gateway DynamoDB Bedrock EventBridge Webhook evolua.io
None
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io
None
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io Authoriser
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io Authoriser
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io Authoriser
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io Authoriser
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io Authoriser
Challenges (for an AI code reviewer) Handling sensitive data for
customers
Challenges (for an AI code reviewer) Large fi les. Large
PRs with many fi les. Handling sensitive data for customers
Why Bedrock?
Security
Security Data is encrypted at rest.
www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak
aws.amazon.com/bedrock/faqs
Security Data is encrypted at rest. Inputs & Outputs are
not shared with model providers. Inputs & Outputs are not used to train other models.
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io Authoriser Fallback
Primary
privacy.anthropic.com/en/articles/7996885-how-do-you-use-personal-data-in-model-training
Serverless
Serverless Usage-based AND provisioned throughput pricing
None
None
1M Input Tokens 1M Output Tokens $0.14 v3 r1 $0.28
$0.55 $2.19 Sonnet $3.75 $15.0 Haiku $0.80 $4.00
Very cost ef fi cient!
Very cost ef fi cient! Data is stored in China.
Very cost ef fi cient! Data is stored in China.
Data might be used to train other models.
www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak
Very cost ef fi cient! Data is stored in China.
Data might be used to train other models. Operationally immature.
None
No token-based pricing yet
No token-based pricing yet “GPU-based instance type like ml.p5e.48xlarge is
recommended”
ml.p5e.48xlarge 💰💰💰💰💰💰💰💰💰💰 💰💰💰💰💰💰💰💰💰💰 💰💰💰💰💰💰💰💰💰💰 💰💰💰💰💰💰💰💰💰💰 💰💰💰💰💰💰💰💰
Other capabilities Guardrails Knowledge base (managed RAG) Agents Cross-region inference
Model evaluations
None
None
None
API Gateway DynamoDB Bedrock EventBridge Webhook AppSync evolua.io Authoriser Fallback
Primary
Lessons
Webhook
Webhook Analyse changes
Webhook Analyse changes Feedback
Condensed view…
None
Lambda timed out after 15 mins
Succeeded on automatic retry
Webhook Analyse changes Feedback LLM limits GitHub limits AWS limits
Lesson: AI is 10% of the problem
None
Reasoning ability
Context window Max response tokens API rate limit Reasoning ability
Context window Max response tokens API rate limit Reasoning ability
Cost Performance
Context window Max response tokens API rate limit Reasoning ability
Cost Performance Important selection criteria for LLMs
Doing cool AI stuff! Working around AI limits
Doing cool AI stuff! Working around AI limits Stop playing
with my bowl…
Context window Max response tokens API rate limit Reasoning ability
Cost Performance
Claude 3.5 Sonnet’s default throughput is 50 per minute
Claude 3.5 Sonnet’s default throughput is 50 per minute Can
be raised to 1,000 per minute
Claude 3.5 Sonnet’s default throughput is 50 per minute Can
be raised to 1,000 per minute Bedrock has cross- region inference
Mitigate API rate limit Raise account limits. Use Bedrock cross-region
inference.
Mitigate API rate limit Raise account limits. Use Bedrock cross-region
inference. Limit no. of parallel requests per PR.
Mitigate API rate limit Raise account limits. Use Bedrock cross-region
inference. Limit no. of parallel requests per PR. Fallback to Anthropic & less powerful models (Claude 3 Sonnet, Claude 3.5 Haiku)
Future work: incorporate other models (Nova, DeepSeek, etc.)
Future work: incorporate other models (Nova, DeepSeek, etc.) Also good
for cost control!
Lesson: LLMs are still quite expensive
None
Dif fi cult to build a sustainable and competitive business
Cost control Only analyse changed lines.
Cost control Only analyse changed lines. Good for cost control
Good for UX
Cost control Only analyse changed lines. Limit free users to
few PRs per month.
API Gateway DynamoDB Bedrock EventBridge Webhook
API Gateway DynamoDB Bedrock EventBridge Webhook Built-in retries & DLQ
Lambda timed out after 15 mins
Lambda timed out after 15 mins Reprocess fi les on
retry…
Lambda timed out after 15 mins Reprocess fi les on
retry… Duplicated side- effects (e.g. Github comments)
Cost control Only analyse changed lines. Limit free users to
few PRs per month. Use checkpoints to avoid re-processing fi les on retries
const issues = await executeIdempotently( `${event-id}-${filename}-analyze`, () => analyzeFile(file) );
... await executeIdempotently( `${event-id}-${filename}-add-gh-comment`, () => addReviewComment(filename, comment) );
Webhook Analyse changes Feedback Why not Step Functions?
Webhook Analyse changes Feedback Why not Step Functions? Checkpoints is
just easier 🤷
Lesson: Latency is a challenge
Models take 10s of seconds to analyse each fi le
Wasted CPU cycles in Lambda
Future work: try other models
Future work: make use of these CPU cycles
Lesson: Be ware of hallucinations
“Give me JSON in this format”
None
“Give me JSON in this format” “Nope!”
None
Non-existent codes, invalid URLs
Non-existent line numbers
Future works
Go to evolua.io to try it out. We’d love your
feedback!
Questions?