AWS Builder Center
Weekend Showcase Challenge: DevLog AI - Turning Raw Commits into Developer Stories

Weekend Showcase Challenge: DevLog AI - Turning Raw Commits into Developer Stories

An automated developer tool powered by Amazon Bedrock and AWS Lambda that turns raw git commit logs and pull requests into polished release notes, changelogs, and technical devlogs in seconds.

Weekend Showcase Challenge: DevLog AI



Every release, the same twenty minutes. Scroll back through  git log --oneline , squint at forty commit subjects written for your future self at 1am, and try to turn them into something a user would actually read. The information is all there. It is just in the wrong shape, written in the wrong voice, for the wrong audience.
DevLog AI is the thing I kept wishing existed while doing that by hand. Paste a git log — or just type  owner/repo  and let it fetch the commits itself — and it gives you back the document you actually need: release notes, a Keep a Changelog entry, an announcement post, or a summary for your team. It streams the result back token by token, and it costs a fraction of a cent per run.
The landing page: four formats, a paste-or-import workbench, and a yellow banner that names which engine is about to answer.

What it does

The core insight is that a commit log is one input with several useful readings. The same twelve commits are a changelog for maintainers, release notes for users, an announcement for the blog, and a "what we shipped" summary for a standup. They are four different documents, not four formats of the same document, so each one gets its own system prompt with its own structure and voice.
Everything else exists to remove friction around that:
  • GitHub import. Every demo used to start with "switch to a terminal and run git log". Now you type  aws/aws-cdk  and it pulls up to 200 commit subjects from the public API.
  • Token streaming. The Lambda runs in  RESPONSE_STREAM  invoke mode, so text appears as it is written rather than after a ten-second pause on a blank panel.
  • Share links. Save a devlog to DynamoDB, get a short URL, drop it in a PR description. Records expire on a TTL so the table stays free-tier sized.
Twelve raw commits in, categorized release notes out. Note that the merge commit and the dependency bump were dropped, and the Conventional Commit scopes became bold prefixes.



Architecture

Five AWS services, no servers, and everything inside the Free Tier.
A flowchart detailing the DevLog AI serverless architecture on AWS, highlighting the connections between a developer client, an AWS Lambda API, Amazon Bedrock for AI generation, Amazon S3 for static site hosting, Amazon DynamoDB for shareable links, and the GitHub REST API.

One Lambda serves the entire application.  /api/*  hits the handlers; every other path is the exported Next.js site, read out of a private S3 bucket.
AWS services used:
Service
Role
AWS Lambda
The whole app. Node 22, arm64, Function URL with  RESPONSE_STREAM .
Amazon Bedrock
Generation, via the Converse API.
Amazon S3
Private bucket holding the Next.js static export.
Amazon DynamoDB
Share links, on-demand billing, TTL expiry.
Amazon CloudFront
Optional edge cache in front (see below).
IAM / CloudWatch Logs
Least-privilege execution role and observability.
The decision I am happiest with is putting the site and the API behind the same Lambda. It started as a workaround and turned out to be the better design: the site and the API share an origin, so there is no CORS layer, no second certificate, and no third service to keep in sync. The bucket stays completely private and is read only by the function's execution role — which is granted  s3:GetObject  and deliberately not  s3:ListBucket .
That one omission caused my favourite bug of the weekend. Every 404 was coming back as a 500. Without  ListBucket , S3 will not confirm that a key is absent to a caller who cannot list — so a missing object returns  403 AccessDenied , not  404 NoSuchKey . My handler was only treating  NoSuchKey  as "missing". The fix was four lines; finding it took an hour.
FLOW DIAGRAM:
A flowchart detailing the backend logic for the generation API. It shows a validation step for the commit log length, followed by a fallback sequence of Amazon Bedrock models ordered by cost (Nova Lite, Nova Micro, Claude 3 Haiku, and Claude Haiku 4.5). If all models fail, it defaults to a local rule-based Conventional Commits parser before streaming the results to the client.




How I built it

The Converse API instead of InvokeModel.  InvokeModel  requires you to know each vendor's request and response shape — Anthropic's  anthropic_version  and content blocks look nothing like Amazon Nova's.  ConverseStream  normalizes both, so Claude and Nova run down exactly the same code path and switching models is a one-line change in a registry. This turned out to matter enormously, because it made the next decision cheap.
A model cascade, ordered cheapest first. The app does not depend on one model being available. It walks a list, and the first model that answers wins.
Real US East on-demand rates, pulled from the AWS Price List API, shown at the point where you choose.
I originally defaulted to Claude Haiku 4.5 and then actually looked at the numbers: $1.10/$5.50 per million tokens, versus $0.06/$0.24 for Amazon Nova Lite. Roughly eighteen times the price for a task that is reshaping text somebody already wrote. Nova Lite became the default. More importantly, I ordered the whole cascade cheapest-first, so falling through an unavailable model steps up in price and you can never be silently escalated into the expensive one. A fifty-commit release note costs well under a tenth of a cent, and the app prints the real cost of each run from the token counts Bedrock reports.
The cascade, and what happens when every model declines.
An honest fallback. If no model will answer, a deterministic Conventional Commits parser renders the same four formats — classifying by type prefix, falling back to keyword heuristics, dropping merge commits and version bumps. It is not a model impersonation. Rule-based output is always badged as such, and a banner names the reason Bedrock was skipped. A showcase app that returns an error page is not a showcase app, but a fallback that quietly passes itself off as AI is worse than an error page.
Screenshots:
The main dashboard of DevLog AI featuring a retro-modern grid layout, format selection options for Release notes, Changelog, Announcement post, and Team summary, alongside a raw Git log input box and generation controls.

A horizontal selector showing the model fallback sequence from cheapest to most capable, including Amazon Nova Lite, Nova Micro, Claude 3 Haiku, and Claude Haiku 4.5, along with estimated per-million token costs.

The preview pane of DevLog AI showing neatly categorized release notes divided into Breaking Changes, Security, New Features, Bug Fixes, Performance, Removed, Documentation, and Internal updates.




The three walls

The real story of this weekend is that I hit three AWS account-level limits that no amount of code could fix, and the app is shaped by how it handles each one.
Bedrock returned  ValidationException: Operation not allowed  for every model, in every region. I spent a while assuming I had the model IDs or the inference profiles wrong. I did not.  aws bedrock get-foundation-model-availability  reported  authorizationStatus: NOT_AUTHORIZED  on everything, while  iam simulate-principal-policy  cheerfully confirmed  bedrock:InvokeModel  was allowed. It is an account entitlement, granted through a support case, and completely invisible from IAM. That is what the rule-based engine is for. Rather than hide it, the app explains it — and when access lands, it switches back to Bedrock on the next request with no redeploy.
CloudFront refused to create a distribution:  AccessDenied: Your account must be verified . So the deploy script detects that, skips CloudFront, and serves the site directly from the Lambda Function URL — which is already HTTPS with a valid certificate. Re-running the script after verification picks CloudFront up with no code change, and the architecture panel in the app reads the real state from  /api/health  rather than claiming a service that is not there.
The Function URL itself returned 403 even with  AuthType: NONE  and a correct resource policy. I diagnosed it by diffing against an older function in the same account that did work: it also carried a  lambda:InvokeFunction  grant, not just  InvokeFunctionUrl . Adding that fixed it immediately.
Building for graceful degradation was not on my plan for the weekend. It became the most interesting part of the project.



What I learned across the summer

In July I built Postcards from Bengaluru — a creative app that turned my city into generated postcards, two a day, sitting in S3. In August I built Daily Auteur, an autonomous cinematographer: a Lambda that wakes on a schedule, reads the actual Bengaluru weather, invents a film still with a title, a logline, a genre, a colour palette and a lens choice, renders it, critiques its own output with a score, and publishes the result. Its commit description still reads "reason, render, critique, publish".
The thread connecting all three is that the AI part kept getting smaller. Postcards was mostly a prompt. Daily Auteur was a prompt plus a loop plus memory in DynamoDB. DevLog AI is maybe fifteen percent model code, and the rest is a cascade, a fallback, an error taxonomy, an idempotent deploy script, and a UI that tells the truth about which engine answered.
The other lesson is more practical: build the non-AI path first. Daily Auteur has been silently falling back to on-box logic for over a week because of the same Bedrock block, and I only discovered it while debugging this project. It never told anyone. DevLog AI tells you, in a yellow box, at the top of the page. That difference is the whole summer in one sentence.



Try it

Paste your own  git log --oneline , or type any public repo and hit Generate. The whole deployment is one script —  npm run deploy  — using nothing but the AWS CLI, and it will not touch anything in your account that it did not create.



A builder who inspired me

Big thanks to @harshendram — still a student, already a community builder, and the person whose projects I kept coming back to this summer. They are aimed at problems people actually have, not at looking clever in a demo, and his dev notes are the kind you open at 1am when something is on fire and they actually help you find it
Thanks to the AWS Builder Center team for running the Summer Build Series. See you in the New to AWS series next month.


Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article